Welcome to Draper's Satellite Image Chronology competition! This one is a little different than our standard fare here on Kaggle. We expect this to be a challenging task for machine and man alike. As such, we are running the competition under rules that reward the best combination of analyst(s) + machine.
\n\nGood luck!
", "messages": [ { "id": 117537, "postDate": "2016-04-29T14:54:04.383Z", "content": "[quote=SecondPlan;117522]
\n\nI am surprised that Kaggle hosts a lot more image related competitions than before.
\n\n[/quote]
\n\nDeep learning/convolutional nets are trending.
Previous Kaggle image contests saw the state-of-the-art being beaten, or have their solutions adapted for production.
Structured data competitions become more-and-more ensembling wars.
[quote=inversion;117580]
\n\nI'm assuming this means we can hand label the test images.
\n\n[/quote]
\n\nCorrect. We think the best approach is going to be some combination of man + machine. Make sure to document what you do either way!
", "rawMarkdown": "[quote=inversion;117580]\r\n\r\nI'm assuming this means we can hand label the **test** images.\r\n\r\n[/quote]\r\n\r\nCorrect. We think the best approach is going to be some combination of man + machine. Make sure to document what you do either way!", "votes": 5 }, { "id": 117540, "postDate": "2016-04-29T15:11:05.387Z", "content": "Just for those who missed it when 'reading' the rules, hand labelling is allowed in this competiton:
\n\n\n\n\nGiven the unique nature of this competition design, please note a\n couple key changes to Kaggle's standard rules, listed below:
\n \nHand labeling and human prediction are permitted. Winners must deliver\n documentation that clearly outlines the manual and automatic steps\n required to create the winning submission. Use of external data is\n permitted.
\n
This should make it rather interesting...
", "rawMarkdown": "Just for those who missed it when 'reading' the rules, hand labelling **is** allowed in this competiton:\r\n\r\n> Given the unique nature of this competition design, please note a\r\n> couple key changes to Kaggle's standard rules, listed below: \r\n> \r\n> Hand labeling and human prediction are permitted. Winners must deliver\r\n> documentation that clearly outlines the manual and automatic steps\r\n> required to create the winning submission. Use of external data is\r\n> permitted.\r\n\r\nThis should make it rather interesting...", "votes": 6 }, { "id": 117529, "postDate": "2016-04-29T14:35:32.043Z", "content": "[quote=SecondPlan;117522]
\n\nI am surprised that Kaggle hosts a lot more image related competitions than before.
\n\n[/quote]
\n\nYeah, me too, which is annoying because image competitions are not my strong suit. I also have a feeling this will be a very difficult competition, as I can't see anything in the images that would suggest when it was taken.
\n\nIt also seems a little weird that the organisers are posting this as a satellite image chronology competition - one would think that satellites have some sort of clock in them to tell when the photo was taken...
", "rawMarkdown": "[quote=SecondPlan;117522]\r\n\r\nI am surprised that Kaggle hosts a lot more image related competitions than before. \r\n\r\n[/quote]\r\n\r\nYeah, me too, which is annoying because image competitions are not my strong suit. I also have a feeling this will be a very difficult competition, as I can't see anything in the images that would suggest when it was taken.\r\n\r\nIt also seems a little weird that the organisers are posting this as a satellite image chronology competition - one would think that satellites have some sort of clock in them to tell when the photo was taken...", "votes": 4 }, { "id": 118531, "postDate": "2016-05-04T05:45:51.087Z", "content": "I agree that this feels like a strange competition (for all reasons listed by the rest of you). For those lamenting not having a titanX, I would not go out and buy one for this competition...given that it's a hard task for a human it would probably be a difficult CNN problem unless you come up with some clever pipeline that somehow isolates important areas...not saying it's impossible but cleverness may win out.
\n\nEveryone has their theories on what is being asked here...
\n\nIt is NOT to improve Draper's space craft landing systems!
\n\nMy pet theory is that it is an intelligence(spy) exercise. What "they" have is many sources of images but they are actually pictures taken of images or computer screens, etc.. The agent(s)(human or otherwise) taking the pictures/video/photocopies/screen-grabs cannot say much about the time stamps, etc.. So "they" are left with a ton of images and a need to time sequence them.
\n\nHa! Like I said, we all have an opinion.
", "rawMarkdown": "I agree that this feels like a strange competition (for all reasons listed by the rest of you). For those lamenting not having a titanX, I would not go out and buy one for this competition...given that it's a hard task for a human it would probably be a difficult CNN problem unless you come up with some clever pipeline that somehow isolates important areas...not saying it's impossible but cleverness may win out.\r\n\r\nEveryone has their theories on what is being asked here...\r\n\r\nIt is NOT to improve Draper's space craft landing systems! \r\n\r\nMy pet theory is that it is an intelligence(spy) exercise. What \"they\" have is many sources of images but they are actually pictures taken of images or computer screens, etc.. The agent(s)(human or otherwise) taking the pictures/video/photocopies/screen-grabs cannot say much about the time stamps, etc.. So \"they\" are left with a ton of images and a need to time sequence them. \r\n\r\nHa! Like I said, we all have an opinion.", "votes": 3 }, { "id": 117876, "postDate": "2016-05-01T10:13:23.177Z", "content": "[quote===>(AL)<==;117674]
\n\nThis all is just feels weird. There is something going on that we are not told ;)
\n\n[/quote]
\n\nI don't see how is this important. We can keep guessing an organizers motivation, for example they may regularly get access to an image series from some third party source, they can not access timestamp data for these images, and they want to be able to put them in series. Or it may be something completely different, but the secrecy doesn't affect very clearly defined task.
", "rawMarkdown": "[quote===>(AL)<==;117674]\r\n\r\nThis all is just feels weird. There is something going on that we are not told ;)\r\n\r\n[/quote]\r\n\r\nI don't see how is this important. We can keep guessing an organizers motivation, for example they may regularly get access to an image series from some third party source, they can not access timestamp data for these images, and they want to be able to put them in series. Or it may be something completely different, but the secrecy doesn't affect very clearly defined task.", "votes": 3 }, { "id": 117637, "postDate": "2016-04-29T22:56:06.710Z", "content": "[quote===>(AL)<==;117635]
\n\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)
\n\n[/quote]
\n\nI wouldn't call randomly guessing the order a productive idea ;)
", "rawMarkdown": "[quote===>(AL)<==;117635]\r\n\r\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)\r\n\r\n[/quote]\r\n\r\nI wouldn't call randomly guessing the order a productive idea ;)", "votes": 3 }, { "id": 117635, "postDate": "2016-04-29T22:42:41.860Z", "content": "[quote=SecondPlan;117611]
\n\n[quote=Triskelion;117537]
\n\n[quote=SecondPlan;117522]
\n\nI am surprised that Kaggle hosts a lot more image related competitions than before.
\n\n[/quote]
\n\nDeep learning/convolutional nets are trending.
Previous Kaggle image contests saw the state-of-the-art being beaten, or have their solutions adapted for production.
Structured data competitions become more-and-more ensembling wars.
[/quote]
\n\nSo Kaggle 2.0 is coming, where ideas are favored than crazing ensemble. But to do CNN on this task, I think you need a good GPU
\n\n[/quote]
\n\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)
", "rawMarkdown": "[quote=SecondPlan;117611]\r\n\r\n[quote=Triskelion;117537]\r\n\r\n[quote=SecondPlan;117522]\r\n\r\nI am surprised that Kaggle hosts a lot more image related competitions than before. \r\n\r\n[/quote]\r\n\r\n* Deep learning/convolutional nets are trending. \r\n\r\n* Previous Kaggle image contests saw the state-of-the-art being beaten, or have their solutions adapted for production. \r\n\r\n* Structured data competitions become more-and-more ensembling wars.\r\n\r\n[/quote]\r\n\r\nSo Kaggle 2.0 is coming, where ideas are favored than crazing ensemble. But to do CNN on this task, I think you need a good GPU\r\n\r\n[/quote]\r\n\r\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)", "votes": 3 }, { "id": 117601, "postDate": "2016-04-29T19:58:29.257Z", "content": "[quote=inversion;117591]
\n\nIf I could poke at the rules a bit . . .
\n\nIn previous contests, it really didn't matter if someone cropped train images, using say, Photoshop, because the test images were never touched.
\n\nBut, what if Photoshop is part of the workflow for test images during this contest? Is that a no-go? Or does one need to do the work in GIMP. What if the functionality is not unique to Photoshop? For example, cropping is pretty standard functionality in all image processing software. Will someone be prize-eligible if they cropped in GIMP, but not if they cropped in Photoshop?
\n\nI supposed there's no way to tell, so maybe it's moot. Unless, of course, someone uses Photoshop functionality that is unavailable in open-source image processing software.
\n\n[/quote]
\n\nWe interpret the rules to favor ideas over technicalities. Cropping is so basic an idea it's like like addition or subtraction; we don't care whether you do it in python, your TI-83, or an abacus. The thing you need to avoid is using closed-source, proprietary magic.
", "rawMarkdown": "[quote=inversion;117591]\r\n\r\nIf I could poke at the rules a bit . . .\r\n\r\nIn previous contests, it really didn't matter if someone cropped train images, using say, Photoshop, because the test images were never touched.\r\n\r\nBut, what if Photoshop is part of the workflow for test images during this contest? Is that a no-go? Or does one need to do the work in GIMP. What if the functionality is not unique to Photoshop? For example, cropping is pretty standard functionality in all image processing software. Will someone be prize-eligible if they cropped in GIMP, but not if they cropped in Photoshop?\r\n\r\nI supposed there's no way to tell, so maybe it's moot. Unless, of course, someone uses Photoshop functionality that is unavailable in open-source image processing software.\r\n\r\n[/quote]\r\n\r\nWe interpret the rules to favor ideas over technicalities. Cropping is so basic an idea it's like like addition or subtraction; we don't care whether you do it in python, your TI-83, or an abacus. The thing you need to avoid is using closed-source, proprietary magic.", "votes": 3 }, { "id": 117580, "postDate": "2016-04-29T18:23:54.600Z", "content": "Hand labeling and human prediction are permitted.
\n\nHand labeling has always been permitted, on the training data. I'm assuming this means we can hand label the test images.
", "rawMarkdown": "_Hand labeling and human prediction are permitted._\r\n\r\nHand labeling has always been permitted, on the **training** data. I'm assuming this means we can hand label the **test** images.", "votes": 3 }, { "id": 117535, "postDate": "2016-04-29T14:51:43.300Z", "content": "[quote=anokas;117529]
\n\nIt also seems a little weird that the organisers are posting this as a satellite image chronology competition - one would think that satellites have some sort of clock in them to tell when the photo was taken...
\n\n[/quote]
\n\nhttp://www.draper.com/news/cracking-code-satellite-data
\n\n\n\n\nAssembling a week’s worth of pictures of the same location in proper chronological order requires contestants to create and demonstrate new tools that can be applied to the expected hundreds of terabytes of daily data from the new constellations.
\n
Also, perhaps they have a baseline of internal intelligence analysts, and they want to see if crowd-sourcing can beat this.
", "rawMarkdown": "[quote=anokas;117529]\r\n\r\nIt also seems a little weird that the organisers are posting this as a satellite image chronology competition - one would think that satellites have some sort of clock in them to tell when the photo was taken...\r\n\r\n[/quote]\r\n\r\nhttp://www.draper.com/news/cracking-code-satellite-data\r\n\r\n> Assembling a week’s worth of pictures of the same location in proper chronological order requires contestants to **create and demonstrate new tools** that can be applied to the expected hundreds of terabytes of daily data from the new constellations.\r\n\r\nAlso, perhaps they have a baseline of internal intelligence analysts, and they want to see if crowd-sourcing can beat this.", "votes": 3 }, { "id": 117520, "postDate": "2016-04-29T14:08:44.797Z", "content": "Welcome to Draper's Satellite Image Chronology competition! This one is a little different than our standard fare here on Kaggle. We expect this to be a challenging task for machine and man alike. As such, we are running the competition under rules that reward the best combination of analyst(s) + machine.
\n\nGood luck!
", "rawMarkdown": "Welcome to Draper's Satellite Image Chronology competition! This one is a little different than our standard fare here on Kaggle. We expect this to be a challenging task for machine and man alike. As such, we are running the competition under rules that reward the best combination of analyst(s) + machine.\r\n\r\nGood luck!", "votes": 3 }, { "id": 117674, "postDate": "2016-04-30T06:08:42.650Z", "content": "[quote=Ben Kamphaus;117672]
\n\nAs someone who worked in the satellite imaging industry for several years I find this to be an odd competition. I'm all about exploring data in new ways or crowd sourcing problems to see if you can find new techniques, but this competition's wording seems to imply that:
\n\nSo what's going on here? It's an interesting puzzle, I guess, but most changing time information are at periods that aren't going to necessarily be visible across the dataset:
\n\nBut it's not like this contains the additional band(s) (e.g. near infrared, etc.) you'd want to estimate vegetation cycles with any accuracy. So maybe you could experiment with several images in the test set and do better than random and ... given that hand selection with additional research is ok, how exactly does this result in something that generalizes or pushes the state of the art forward?
\n\nAnyways, I'm gonna do a bit more exploratory analysis and think it over, but this is just a weird feeling contest. Like I'm wondering if there are messages encrypted in the imagery in small noise values that don't impact how the images appear viewed under typical image histogram stretch sensitivities level, and the contest is really about finding and decoding those.
\n\n[/quote]
\n\nMy thoughts exactly. I also spend quite a bit of time with satellite imagery and even thought Satellite Remote Sensing class to master students. The topic says "satellite imagery" but it has nothing to do with it in reality. It gives aerophoto instead of satellite - it's like saying that an automobile and an oax cart is the same thing. The both have wheels! it implies that time-stamping of an image is an important task, but I have never seen a satellite image without a timestamp (and geo-location information, sun elevation, etc, in one way or another)
\n\nI have a feel that the real purpose of this puzzle was not the one we were given. Furthermore, it all can be relatively easily done - images can be co-registered, change detection is at least several decades considered a mature science and can be done easily on co-registered images. Nothing of these activities will really move state of the art forward. Maybe sponsors are after new techniques for object localization and recognition - but why not to say so?
\n\nThere is another weird thing about this contest - the target is not well defined. We are asked to produce the right order for images taken over several days. We can, of course, to find that cars in the residential neighborhood are parked differently, for example. I could have imagined that this could have been about time series, and establishing some novel ways of combining RNN with CNN for example. But this is apparently not the case either. We could have used time effects to order things if images were overlapping with some movement, like cars moving on a highway. Or if images were taken over longer time interval when vegetation changes. But images are taken several days apart, in unknown part of the day, so no real time series either. Train and test images are taken from the different areas in the most part. There is very little overlap between train and test. Hand labeling is allowed.
\n\nThis all is just feels weird. There is something going on that we are not told ;)
", "rawMarkdown": "[quote=Ben Kamphaus;117672]\r\n\r\nAs someone who worked in the satellite imaging industry for several years I find this to be an odd competition. I'm all about exploring data in new ways or crowd sourcing problems to see if you can find new techniques, but this competition's wording seems to imply that:\r\n\r\n* this is the kind of imagery earth observing satellites typically generate for analysts to use (it isn't)\r\n* there hasn't been much innovation regarding time series recently (there's been a ton, just google scholar search for \"satellite time series imagery\")\r\n* that change detection is hard (in the vast majority of cases it's trivial)\r\n* that this is the kind of task people people need to figure out how to do (it's not, satellite ephemeris data with precise time stamps is pretty much universal).\r\n\r\nSo what's going on here? It's an interesting puzzle, I guess, but most changing time information are at periods that aren't going to necessarily be visible across the dataset:\r\n\r\n* vegetation changes fairly minute over day to day periods, but significant over seasons.\r\n* activity around buildings, traffic, etc. changes for time of day and specific day\r\n* the sensor may move or have moved in a particular path that makes sense?\r\n\r\nBut it's not like this contains the additional band(s) (e.g. near infrared, etc.) you'd want to estimate vegetation cycles with any accuracy. So maybe you could experiment with several images in the test set and do better than random and ... given that hand selection with additional research is ok, how exactly does this result in something that generalizes or pushes the state of the art forward?\r\n\r\nAnyways, I'm gonna do a bit more exploratory analysis and think it over, but this is just a weird feeling contest. Like I'm wondering if there are messages encrypted in the imagery in small noise values that don't impact how the images appear viewed under typical image histogram stretch sensitivities level, and the contest is really about finding and decoding those.\r\n\r\n[/quote]\r\n\r\nMy thoughts exactly. I also spend quite a bit of time with satellite imagery and even thought Satellite Remote Sensing class to master students. The topic says \"satellite imagery\" but it has nothing to do with it in reality. It gives aerophoto instead of satellite - it's like saying that an automobile and an oax cart is the same thing. The both have wheels! it implies that time-stamping of an image is an important task, but I have never seen a satellite image without a timestamp (and geo-location information, sun elevation, etc, in one way or another)\r\n\r\nI have a feel that the real purpose of this puzzle was not the one we were given. Furthermore, it all can be relatively easily done - images can be co-registered, change detection is at least several decades considered a mature science and can be done easily on co-registered images. Nothing of these activities will really move state of the art forward. Maybe sponsors are after new techniques for object localization and recognition - but why not to say so?\r\n\r\nThere is another weird thing about this contest - the target is not well defined. We are asked to produce the right order for images taken over several days. We can, of course, to find that cars in the residential neighborhood are parked differently, for example. I could have imagined that this could have been about time series, and establishing some novel ways of combining RNN with CNN for example. But this is apparently not the case either. We could have used time effects to order things if images were overlapping with some movement, like cars moving on a highway. Or if images were taken over longer time interval when vegetation changes. But images are taken several days apart, in unknown part of the day, so no real time series either. Train and test images are taken from the different areas in the most part. There is very little overlap between train and test. Hand labeling is allowed. \r\n\r\nThis all is just feels weird. There is something going on that we are not told ;)", "votes": 2 }, { "id": 118047, "postDate": "2016-05-02T18:00:05.227Z", "content": "[quote===>(AL)<==;117648]
\n\n[quote=anokas;117637]
\n\n[quote===>(AL)<==;117635]
\n\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)
\n\n[/quote]
\n\nI wouldn't call randomly guessing the order a productive idea ;)
\n\n[/quote]
\n\nIf we define productive as giving people high LB standing then it was a productive idea, at least at the time
\n\n[/quote]
\n\nInteresting insight about the motives. Maybe someone wants to see if you guys can detect timestamp tampering or not ?
", "rawMarkdown": "[quote===>(AL)<==;117648]\r\n\r\n[quote=anokas;117637]\r\n\r\n[quote===>(AL)<==;117635]\r\n\r\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)\r\n\r\n[/quote]\r\n\r\nI wouldn't call randomly guessing the order a productive idea ;)\r\n\r\n[/quote]\r\n\r\nIf we define productive as giving people high LB standing then it was a productive idea, at least at the time\r\n\r\n\r\n[/quote]\r\n\r\nInteresting insight about the motives. Maybe someone wants to see if you guys can detect timestamp tampering or not ? ", "votes": 1 }, { "id": 117772, "postDate": "2016-04-30T20:36:30.183Z", "content": "Yes, the metric will give -1 for perfect anticorrelation.
", "rawMarkdown": "Yes, the metric will give -1 for perfect anticorrelation.", "votes": 1 }, { "id": 117648, "postDate": "2016-04-30T01:04:53.280Z", "content": "[quote=anokas;117637]
\n\n[quote===>(AL)<==;117635]
\n\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)
\n\n[/quote]
\n\nI wouldn't call randomly guessing the order a productive idea ;)
\n\n[/quote]
\n\nIf we define productive as giving people high LB standing then it was a productive idea, at least at the time
", "rawMarkdown": "[quote=anokas;117637]\r\n\r\n[quote===>(AL)<==;117635]\r\n\r\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)\r\n\r\n[/quote]\r\n\r\nI wouldn't call randomly guessing the order a productive idea ;)\r\n\r\n[/quote]\r\n\r\nIf we define productive as giving people high LB standing then it was a productive idea, at least at the time\r\n", "votes": 1 }, { "id": 117634, "postDate": "2016-04-29T22:39:55.517Z", "content": "[quote=William Cukierski;117603]
\n\n[quote=Jmh1;117593]
\n\nJust to clarify, each picture of a set was taken on a different day. They are not the same day, but at different times or angles?
\n\n[/quote]
\n\nCorrect
\n\n[/quote]
\n\nThis wasn't clear from the description that each image is taken on a different day? So there is no rapid-shot sequences within few minutes/seconds from the same day? All images from a given set are taken at different days and possibly from different time of the day, correct? Are they within a short time frame from each other - like days, or some may be month or years apart?
", "rawMarkdown": "[quote=William Cukierski;117603]\r\n\r\n[quote=Jmh1;117593]\r\n\r\nJust to clarify, each picture of a set was taken on a different day. They are not the same day, but at different times or angles? \r\n\r\n[/quote]\r\n\r\nCorrect\r\n\r\n[/quote]\r\n\r\nThis wasn't clear from the description that each image is taken on a **different** day? So there is no rapid-shot sequences within few minutes/seconds from the same day? **All** images from a given set are taken at **different** days and possibly from different time of the day, correct? Are they within a short time frame from each other - like days, or some may be month or years apart?", "votes": 1 }, { "id": 117611, "postDate": "2016-04-29T20:33:57.537Z", "content": "[quote=Triskelion;117537]
\n\n[quote=SecondPlan;117522]
\n\nI am surprised that Kaggle hosts a lot more image related competitions than before.
\n\n[/quote]
\n\nDeep learning/convolutional nets are trending.
Previous Kaggle image contests saw the state-of-the-art being beaten, or have their solutions adapted for production.
Structured data competitions become more-and-more ensembling wars.
[/quote]
\n\nSo Kaggle 2.0 is coming, where ideas are favored than crazing ensemble. But to do CNN on this task, I think you need a good GPU
", "rawMarkdown": "[quote=Triskelion;117537]\r\n\r\n[quote=SecondPlan;117522]\r\n\r\nI am surprised that Kaggle hosts a lot more image related competitions than before. \r\n\r\n[/quote]\r\n\r\n* Deep learning/convolutional nets are trending. \r\n\r\n* Previous Kaggle image contests saw the state-of-the-art being beaten, or have their solutions adapted for production. \r\n\r\n* Structured data competitions become more-and-more ensembling wars.\r\n\r\n[/quote]\r\n\r\nSo Kaggle 2.0 is coming, where ideas are favored than crazing ensemble. But to do CNN on this task, I think you need a good GPU", "votes": 1 }, { "id": 117593, "postDate": "2016-04-29T19:16:17.037Z", "content": "Just to clarify, each picture of a set was taken on a different day? They are not the same day, but at different times or angles?
", "rawMarkdown": "Just to clarify, each picture of a set was taken on a different day? They are not the same day, but at different times or angles? ", "votes": 1 }, { "id": 117591, "postDate": "2016-04-29T18:59:16.620Z", "content": "If I could poke at the rules a bit . . .
\n\nIn previous contests, it really didn't matter if someone cropped train images, using say, Photoshop, because the test images were never touched.
\n\nBut, what if Photoshop is part of the workflow for test images during this contest? Is that a no-go? Or does one need to do the work in GIMP. What if the functionality is not unique to Photoshop? For example, cropping is pretty standard functionality in all image processing software. Will someone be prize-eligible if they cropped in GIMP, but not if they cropped in Photoshop?
\n\nI supposed there's no way to tell, so maybe it's moot. Unless, of course, someone uses Photoshop functionality that is unavailable in open-source image processing software.
", "rawMarkdown": "If I could poke at the rules a bit . . .\r\n\r\nIn previous contests, it really didn't matter if someone cropped train images, using say, Photoshop, because the test images were never touched.\r\n\r\nBut, what if Photoshop is part of the workflow for test images during this contest? Is that a no-go? Or does one need to do the work in GIMP. What if the functionality is not unique to Photoshop? For example, cropping is pretty standard functionality in all image processing software. Will someone be prize-eligible if they cropped in GIMP, but not if they cropped in Photoshop?\r\n\r\nI supposed there's no way to tell, so maybe it's moot. Unless, of course, someone uses Photoshop functionality that is unavailable in open-source image processing software.", "votes": 1 }, { "id": 117553, "postDate": "2016-04-29T16:18:15.820Z", "content": "This competition reminds me of the skynet in the movie of Terminator. Fellow Kaggler, are you ready to save the human race?
\n\nThe trick of this competition is identify the mobile feature, such as car, from fixed structure.
\n\nThe shadow of the building is very interesting. Are the pictures taken exactly the same time during the five day period?
", "rawMarkdown": "This competition reminds me of the skynet in the movie of Terminator. Fellow Kaggler, are you ready to save the human race?\r\n\r\n\r\nThe trick of this competition is identify the mobile feature, such as car, from fixed structure.\r\n\r\nThe shadow of the building is very interesting. Are the pictures taken exactly the same time during the five day period? ", "votes": 1 }, { "id": 117560, "postDate": "2016-04-29T16:40:27.607Z", "content": "[quote=Birchwood;117553]
\n\nThe shadow of the building is very interesting. Are the pictures taken exactly the same time during the five day period?
\n\n[/quote]
\n\nFrom the data page:
\n\nEach image in a set was taken on a different day (but not necessarily at the same time each day).
", "rawMarkdown": "[quote=Birchwood;117553]\r\n\r\nThe shadow of the building is very interesting. Are the pictures taken exactly the same time during the five day period? \r\n\r\n[/quote]\r\n\r\nFrom the data page:\r\n\r\nEach image in a set was taken on a different day (but not necessarily at the same time each day).", "votes": 2 }, { "id": 117524, "postDate": "2016-04-29T14:22:47.530Z", "content": "I convinced that I must buy a titan x.
", "rawMarkdown": "I convinced that I must buy a titan x.", "votes": 2 }, { "id": 117766, "postDate": "2016-04-30T19:56:58.827Z", "content": "I have a question about metric used for leaderboard. If correct answer is 1 2 3 4 5, but I output 5 4 3 2 1. Will it give -1 as score?
\n\n print(spearmanr([1,2,3,4,5], [1,2,3,4,5]))\n print(spearmanr([1,2,3,4,5], [1,2,4,3,5]))\n print(spearmanr([1,2,3,4,5], [5,1,2,3,4]))\n print(spearmanr([1,2,3,4,5], [5,4,3,2,1]))\n print(spearmanr([1,2,3,4,5], [5,4,2,3,1]))\n\n(1.0, 0.0)\n(0.9, 0.037)\n(0.0, 1.0)\n(-1.0, 0.0)\n(-0.9, 0.037)\nI think knowing the right order is actually better than random guess producing the better score.
", "rawMarkdown": "I have a question about metric used for leaderboard. If correct answer is 1 2 3 4 5, but I output 5 4 3 2 1. Will it give -1 as score?\r\n\r\n print(spearmanr([1,2,3,4,5], [1,2,3,4,5]))\r\n print(spearmanr([1,2,3,4,5], [1,2,4,3,5]))\r\n print(spearmanr([1,2,3,4,5], [5,1,2,3,4]))\r\n print(spearmanr([1,2,3,4,5], [5,4,3,2,1]))\r\n print(spearmanr([1,2,3,4,5], [5,4,2,3,1]))\r\n\r\n (1.0, 0.0)\r\n (0.9, 0.037)\r\n (0.0, 1.0)\r\n (-1.0, 0.0)\r\n (-0.9, 0.037)\r\n\r\nI think knowing the right order is actually better than random guess producing the better score.\r\n" }, { "id": 117576, "postDate": "2016-04-29T18:16:11.653Z", "content": "If I understand this correctly, all the images within one set were taken on the same day ?
", "rawMarkdown": "If I understand this correctly, all the images within one set were taken on the same day ?\r\n", "votes": -2 }, { "id": 119456, "postDate": "2016-05-10T12:25:46.473Z", "content": "I also like the spy theory. But according to challenge description the objective is to detect subtle day to day changes in satellite images. So it's not about detecting time but about detecting subtle changes and patterns. It looks more like an open ended research problem rather a practical one. The output of the challenge will be novel tools and techniques to analyze such aerial images, and maybe some patterns we discover analyzing these images. I agree that this would have military/intelligence applications.
", "rawMarkdown": "I also like the spy theory. But according to challenge description the objective is to detect subtle day to day changes in satellite images. So it's not about detecting time but about detecting subtle changes and patterns. It looks more like an open ended research problem rather a practical one. The output of the challenge will be novel tools and techniques to analyze such aerial images, and maybe some patterns we discover analyzing these images. I agree that this would have military/intelligence applications." }, { "id": 119410, "postDate": "2016-05-10T02:02:45.087Z", "content": "Can someone point to the real objective behind this exercise (Just for motivation)\n+1 for spy theory though!\nThanks!
", "rawMarkdown": "Can someone point to the real objective behind this exercise (Just for motivation)\r\n+1 for spy theory though!\r\nThanks!" }, { "id": 119248, "postDate": "2016-05-08T12:17:54.607Z", "content": "Have the images been post-processed in any way after acquisition (besides an obvious removal of the metadata)? Rotation, crop, contrast adjustments, anything like that?
", "rawMarkdown": "Have the images been post-processed in any way after acquisition (besides an obvious removal of the metadata)? Rotation, crop, contrast adjustments, anything like that?" }, { "id": 119209, "postDate": "2016-05-08T03:43:50.077Z", "content": "[quote=zero zero;118531]
\n\nI agree that this feels like a strange competition (for all reasons listed by the rest of you). For those lamenting not having a titanX, I would not go out and buy one for this competition...given that it's a hard task for a human it would probably be a difficult CNN problem unless you come up with some clever pipeline that somehow isolates important areas...not saying it's impossible but cleverness may win out.
\n\nEveryone has their theories on what is being asked here...
\n\nIt is NOT to improve Draper's space craft landing systems!
\n\nMy pet theory is that it is an intelligence(spy) exercise. What "they" have is many sources of images but they are actually pictures taken of images or computer screens, etc.. The agent(s)(human or otherwise) taking the pictures/video/photocopies/screen-grabs cannot say much about the time stamps, etc.. So "they" are left with a ton of images and a need to time sequence them.
\n\nHa! Like I said, we all have an opinion.
\n\n[/quote]
\n\nI like this spy story!
", "rawMarkdown": "[quote=zero zero;118531]\r\n\r\nI agree that this feels like a strange competition (for all reasons listed by the rest of you). For those lamenting not having a titanX, I would not go out and buy one for this competition...given that it's a hard task for a human it would probably be a difficult CNN problem unless you come up with some clever pipeline that somehow isolates important areas...not saying it's impossible but cleverness may win out.\r\n\r\nEveryone has their theories on what is being asked here...\r\n\r\nIt is NOT to improve Draper's space craft landing systems! \r\n\r\nMy pet theory is that it is an intelligence(spy) exercise. What \"they\" have is many sources of images but they are actually pictures taken of images or computer screens, etc.. The agent(s)(human or otherwise) taking the pictures/video/photocopies/screen-grabs cannot say much about the time stamps, etc.. So \"they\" are left with a ton of images and a need to time sequence them. \r\n\r\nHa! Like I said, we all have an opinion.\r\n\r\n[/quote]\r\n\r\nI like this spy story!" }, { "id": 118134, "postDate": "2016-05-02T22:11:48.020Z", "content": "Just a thought - I wonder if the data is meant to be similar to some sort of drone image set. It seems closer to that than any usual kind of satellite imagery or rigorously-taken aerial photography.
", "rawMarkdown": "Just a thought - I wonder if the data is meant to be similar to some sort of drone image set. It seems closer to that than any usual kind of satellite imagery or rigorously-taken aerial photography." }, { "id": 117756, "postDate": "2016-04-30T17:58:24.030Z", "content": "[quote=I can do the splits NO PROB;117693]
\n\n@AL: The untold story? Have you visited draper.com? ;)
\n\n[/quote]
\n\nI have. I've known them before this ;)
", "rawMarkdown": "[quote=I can do the splits NO PROB;117693]\r\n\r\n@AL: The untold story? Have you visited draper.com? ;)\r\n\r\n[/quote]\r\n\r\nI have. I've known them before this ;)" }, { "id": 117603, "postDate": "2016-04-29T19:59:46.347Z", "content": "[quote=Jmh1;117593]
\n\nJust to clarify, each picture of a set was taken on a different day? They are not the same day, but at different times or angles?
\n\n[/quote]
\n\nCorrect
", "rawMarkdown": "[quote=Jmh1;117593]\r\n\r\nJust to clarify, each picture of a set was taken on a different day? They are not the same day, but at different times or angles? \r\n\r\n[/quote]\r\n\r\nCorrect" }, { "id": 117522, "postDate": "2016-04-29T14:13:30.373Z", "content": "I am surprised that Kaggle hosts a lot more image related competitions than before.
", "rawMarkdown": "I am surprised that Kaggle hosts a lot more image related competitions than before. " }, { "id": 117693, "postDate": "2016-04-30T09:17:25.593Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 117672, "postDate": "2016-04-30T05:21:17.553Z", "rawMarkdown": "", "votes": 9, "isDeleted": true } ], "comments": [ { "id": 117537, "author_name": "Triskelion", "author_url": "", "post_date": "2016-04-29T14:54:04.383000", "content": "[quote=SecondPlan;117522]
\n\nI am surprised that Kaggle hosts a lot more image related competitions than before.
\n\n[/quote]
\n\nDeep learning/convolutional nets are trending.
Previous Kaggle image contests saw the state-of-the-art being beaten, or have their solutions adapted for production.
Structured data competitions become more-and-more ensembling wars.
[quote=inversion;117580]
\n\nI'm assuming this means we can hand label the test images.
\n\n[/quote]
\n\nCorrect. We think the best approach is going to be some combination of man + machine. Make sure to document what you do either way!
", "votes": 5, "replies": [] }, { "id": 117540, "author_name": "anokas", "author_url": "", "post_date": "2016-04-29T15:11:05.387000", "content": "Just for those who missed it when 'reading' the rules, hand labelling is allowed in this competiton:
\n\n\n\n\nGiven the unique nature of this competition design, please note a\n couple key changes to Kaggle's standard rules, listed below:
\n \nHand labeling and human prediction are permitted. Winners must deliver\n documentation that clearly outlines the manual and automatic steps\n required to create the winning submission. Use of external data is\n permitted.
\n
This should make it rather interesting...
", "votes": 6, "replies": [] }, { "id": 117529, "author_name": "anokas", "author_url": "", "post_date": "2016-04-29T14:35:32.043000", "content": "[quote=SecondPlan;117522]
\n\nI am surprised that Kaggle hosts a lot more image related competitions than before.
\n\n[/quote]
\n\nYeah, me too, which is annoying because image competitions are not my strong suit. I also have a feeling this will be a very difficult competition, as I can't see anything in the images that would suggest when it was taken.
\n\nIt also seems a little weird that the organisers are posting this as a satellite image chronology competition - one would think that satellites have some sort of clock in them to tell when the photo was taken...
", "votes": 4, "replies": [] }, { "id": 118531, "author_name": "zero zero", "author_url": "", "post_date": "2016-05-04T05:45:51.087000", "content": "I agree that this feels like a strange competition (for all reasons listed by the rest of you). For those lamenting not having a titanX, I would not go out and buy one for this competition...given that it's a hard task for a human it would probably be a difficult CNN problem unless you come up with some clever pipeline that somehow isolates important areas...not saying it's impossible but cleverness may win out.
\n\nEveryone has their theories on what is being asked here...
\n\nIt is NOT to improve Draper's space craft landing systems!
\n\nMy pet theory is that it is an intelligence(spy) exercise. What "they" have is many sources of images but they are actually pictures taken of images or computer screens, etc.. The agent(s)(human or otherwise) taking the pictures/video/photocopies/screen-grabs cannot say much about the time stamps, etc.. So "they" are left with a ton of images and a need to time sequence them.
\n\nHa! Like I said, we all have an opinion.
", "votes": 3, "replies": [] }, { "id": 117876, "author_name": "itod", "author_url": "", "post_date": "2016-05-01T10:13:23.177000", "content": "[quote===>(AL)<==;117674]
\n\nThis all is just feels weird. There is something going on that we are not told ;)
\n\n[/quote]
\n\nI don't see how is this important. We can keep guessing an organizers motivation, for example they may regularly get access to an image series from some third party source, they can not access timestamp data for these images, and they want to be able to put them in series. Or it may be something completely different, but the secrecy doesn't affect very clearly defined task.
", "votes": 3, "replies": [] }, { "id": 117637, "author_name": "anokas", "author_url": "", "post_date": "2016-04-29T22:56:06.710000", "content": "[quote===>(AL)<==;117635]
\n\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)
\n\n[/quote]
\n\nI wouldn't call randomly guessing the order a productive idea ;)
", "votes": 3, "replies": [] }, { "id": 117635, "author_name": "==>(AL)<==", "author_url": "", "post_date": "2016-04-29T22:42:41.860000", "content": "[quote=SecondPlan;117611]
\n\n[quote=Triskelion;117537]
\n\n[quote=SecondPlan;117522]
\n\nI am surprised that Kaggle hosts a lot more image related competitions than before.
\n\n[/quote]
\n\nDeep learning/convolutional nets are trending.
Previous Kaggle image contests saw the state-of-the-art being beaten, or have their solutions adapted for production.
Structured data competitions become more-and-more ensembling wars.
[/quote]
\n\nSo Kaggle 2.0 is coming, where ideas are favored than crazing ensemble. But to do CNN on this task, I think you need a good GPU
\n\n[/quote]
\n\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)
", "votes": 3, "replies": [] }, { "id": 117601, "author_name": "Will Cukierski", "author_url": "", "post_date": "2016-04-29T19:58:29.257000", "content": "[quote=inversion;117591]
\n\nIf I could poke at the rules a bit . . .
\n\nIn previous contests, it really didn't matter if someone cropped train images, using say, Photoshop, because the test images were never touched.
\n\nBut, what if Photoshop is part of the workflow for test images during this contest? Is that a no-go? Or does one need to do the work in GIMP. What if the functionality is not unique to Photoshop? For example, cropping is pretty standard functionality in all image processing software. Will someone be prize-eligible if they cropped in GIMP, but not if they cropped in Photoshop?
\n\nI supposed there's no way to tell, so maybe it's moot. Unless, of course, someone uses Photoshop functionality that is unavailable in open-source image processing software.
\n\n[/quote]
\n\nWe interpret the rules to favor ideas over technicalities. Cropping is so basic an idea it's like like addition or subtraction; we don't care whether you do it in python, your TI-83, or an abacus. The thing you need to avoid is using closed-source, proprietary magic.
", "votes": 3, "replies": [] }, { "id": 117580, "author_name": "inversion", "author_url": "", "post_date": "2016-04-29T18:23:54.600000", "content": "Hand labeling and human prediction are permitted.
\n\nHand labeling has always been permitted, on the training data. I'm assuming this means we can hand label the test images.
", "votes": 3, "replies": [] }, { "id": 117535, "author_name": "Triskelion", "author_url": "", "post_date": "2016-04-29T14:51:43.300000", "content": "[quote=anokas;117529]
\n\nIt also seems a little weird that the organisers are posting this as a satellite image chronology competition - one would think that satellites have some sort of clock in them to tell when the photo was taken...
\n\n[/quote]
\n\nhttp://www.draper.com/news/cracking-code-satellite-data
\n\n\n\n\nAssembling a week’s worth of pictures of the same location in proper chronological order requires contestants to create and demonstrate new tools that can be applied to the expected hundreds of terabytes of daily data from the new constellations.
\n
Also, perhaps they have a baseline of internal intelligence analysts, and they want to see if crowd-sourcing can beat this.
", "votes": 3, "replies": [] }, { "id": 117674, "author_name": "==>(AL)<==", "author_url": "", "post_date": "2016-04-30T06:08:42.650000", "content": "[quote=Ben Kamphaus;117672]
\n\nAs someone who worked in the satellite imaging industry for several years I find this to be an odd competition. I'm all about exploring data in new ways or crowd sourcing problems to see if you can find new techniques, but this competition's wording seems to imply that:
\n\nSo what's going on here? It's an interesting puzzle, I guess, but most changing time information are at periods that aren't going to necessarily be visible across the dataset:
\n\nBut it's not like this contains the additional band(s) (e.g. near infrared, etc.) you'd want to estimate vegetation cycles with any accuracy. So maybe you could experiment with several images in the test set and do better than random and ... given that hand selection with additional research is ok, how exactly does this result in something that generalizes or pushes the state of the art forward?
\n\nAnyways, I'm gonna do a bit more exploratory analysis and think it over, but this is just a weird feeling contest. Like I'm wondering if there are messages encrypted in the imagery in small noise values that don't impact how the images appear viewed under typical image histogram stretch sensitivities level, and the contest is really about finding and decoding those.
\n\n[/quote]
\n\nMy thoughts exactly. I also spend quite a bit of time with satellite imagery and even thought Satellite Remote Sensing class to master students. The topic says "satellite imagery" but it has nothing to do with it in reality. It gives aerophoto instead of satellite - it's like saying that an automobile and an oax cart is the same thing. The both have wheels! it implies that time-stamping of an image is an important task, but I have never seen a satellite image without a timestamp (and geo-location information, sun elevation, etc, in one way or another)
\n\nI have a feel that the real purpose of this puzzle was not the one we were given. Furthermore, it all can be relatively easily done - images can be co-registered, change detection is at least several decades considered a mature science and can be done easily on co-registered images. Nothing of these activities will really move state of the art forward. Maybe sponsors are after new techniques for object localization and recognition - but why not to say so?
\n\nThere is another weird thing about this contest - the target is not well defined. We are asked to produce the right order for images taken over several days. We can, of course, to find that cars in the residential neighborhood are parked differently, for example. I could have imagined that this could have been about time series, and establishing some novel ways of combining RNN with CNN for example. But this is apparently not the case either. We could have used time effects to order things if images were overlapping with some movement, like cars moving on a highway. Or if images were taken over longer time interval when vegetation changes. But images are taken several days apart, in unknown part of the day, so no real time series either. Train and test images are taken from the different areas in the most part. There is very little overlap between train and test. Hand labeling is allowed.
\n\nThis all is just feels weird. There is something going on that we are not told ;)
", "votes": 2, "replies": [] }, { "id": 118047, "author_name": "Jose Berengueres", "author_url": "", "post_date": "2016-05-02T18:00:05.227000", "content": "[quote===>(AL)<==;117648]
\n\n[quote=anokas;117637]
\n\n[quote===>(AL)<==;117635]
\n\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)
\n\n[/quote]
\n\nI wouldn't call randomly guessing the order a productive idea ;)
\n\n[/quote]
\n\nIf we define productive as giving people high LB standing then it was a productive idea, at least at the time
\n\n[/quote]
\n\nInteresting insight about the motives. Maybe someone wants to see if you guys can detect timestamp tampering or not ?
", "votes": 1, "replies": [] }, { "id": 117772, "author_name": "Will Cukierski", "author_url": "", "post_date": "2016-04-30T20:36:30.183000", "content": "Yes, the metric will give -1 for perfect anticorrelation.
", "votes": 1, "replies": [] }, { "id": 117648, "author_name": "==>(AL)<==", "author_url": "", "post_date": "2016-04-30T01:04:53.280000", "content": "[quote=anokas;117637]
\n\n[quote===>(AL)<==;117635]
\n\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)
\n\n[/quote]
\n\nI wouldn't call randomly guessing the order a productive idea ;)
\n\n[/quote]
\n\nIf we define productive as giving people high LB standing then it was a productive idea, at least at the time
", "votes": 1, "replies": [] }, { "id": 117634, "author_name": "==>(AL)<==", "author_url": "", "post_date": "2016-04-29T22:39:55.517000", "content": "[quote=William Cukierski;117603]
\n\n[quote=Jmh1;117593]
\n\nJust to clarify, each picture of a set was taken on a different day. They are not the same day, but at different times or angles?
\n\n[/quote]
\n\nCorrect
\n\n[/quote]
\n\nThis wasn't clear from the description that each image is taken on a different day? So there is no rapid-shot sequences within few minutes/seconds from the same day? All images from a given set are taken at different days and possibly from different time of the day, correct? Are they within a short time frame from each other - like days, or some may be month or years apart?
", "votes": 1, "replies": [] }, { "id": 117611, "author_name": "SecondPlan", "author_url": "", "post_date": "2016-04-29T20:33:57.537000", "content": "[quote=Triskelion;117537]
\n\n[quote=SecondPlan;117522]
\n\nI am surprised that Kaggle hosts a lot more image related competitions than before.
\n\n[/quote]
\n\nDeep learning/convolutional nets are trending.
Previous Kaggle image contests saw the state-of-the-art being beaten, or have their solutions adapted for production.
Structured data competitions become more-and-more ensembling wars.
[/quote]
\n\nSo Kaggle 2.0 is coming, where ideas are favored than crazing ensemble. But to do CNN on this task, I think you need a good GPU
", "votes": 1, "replies": [] }, { "id": 117593, "author_name": "Jmh1", "author_url": "", "post_date": "2016-04-29T19:16:17.037000", "content": "Just to clarify, each picture of a set was taken on a different day? They are not the same day, but at different times or angles?
", "votes": 1, "replies": [] }, { "id": 117591, "author_name": "inversion", "author_url": "", "post_date": "2016-04-29T18:59:16.620000", "content": "If I could poke at the rules a bit . . .
\n\nIn previous contests, it really didn't matter if someone cropped train images, using say, Photoshop, because the test images were never touched.
\n\nBut, what if Photoshop is part of the workflow for test images during this contest? Is that a no-go? Or does one need to do the work in GIMP. What if the functionality is not unique to Photoshop? For example, cropping is pretty standard functionality in all image processing software. Will someone be prize-eligible if they cropped in GIMP, but not if they cropped in Photoshop?
\n\nI supposed there's no way to tell, so maybe it's moot. Unless, of course, someone uses Photoshop functionality that is unavailable in open-source image processing software.
", "votes": 1, "replies": [] }, { "id": 117553, "author_name": "Jialun He", "author_url": "", "post_date": "2016-04-29T16:18:15.820000", "content": "This competition reminds me of the skynet in the movie of Terminator. Fellow Kaggler, are you ready to save the human race?
\n\nThe trick of this competition is identify the mobile feature, such as car, from fixed structure.
\n\nThe shadow of the building is very interesting. Are the pictures taken exactly the same time during the five day period?
", "votes": 1, "replies": [] }, { "id": 117560, "author_name": "anokas", "author_url": "", "post_date": "2016-04-29T16:40:27.607000", "content": "[quote=Birchwood;117553]
\n\nThe shadow of the building is very interesting. Are the pictures taken exactly the same time during the five day period?
\n\n[/quote]
\n\nFrom the data page:
\n\nEach image in a set was taken on a different day (but not necessarily at the same time each day).
", "votes": 2, "replies": [] }, { "id": 117524, "author_name": "khyh", "author_url": "", "post_date": "2016-04-29T14:22:47.530000", "content": "I convinced that I must buy a titan x.
", "votes": 2, "replies": [] }, { "id": 117766, "author_name": "ZFTurbo", "author_url": "", "post_date": "2016-04-30T19:56:58.827000", "content": "I have a question about metric used for leaderboard. If correct answer is 1 2 3 4 5, but I output 5 4 3 2 1. Will it give -1 as score?
\n\n print(spearmanr([1,2,3,4,5], [1,2,3,4,5]))\n print(spearmanr([1,2,3,4,5], [1,2,4,3,5]))\n print(spearmanr([1,2,3,4,5], [5,1,2,3,4]))\n print(spearmanr([1,2,3,4,5], [5,4,3,2,1]))\n print(spearmanr([1,2,3,4,5], [5,4,2,3,1]))\n\n(1.0, 0.0)\n(0.9, 0.037)\n(0.0, 1.0)\n(-1.0, 0.0)\n(-0.9, 0.037)\nI think knowing the right order is actually better than random guess producing the better score.
", "votes": 0, "replies": [] }, { "id": 117576, "author_name": "Jonathan Chung", "author_url": "", "post_date": "2016-04-29T18:16:11.653000", "content": "If I understand this correctly, all the images within one set were taken on the same day ?
", "votes": -2, "replies": [] }, { "id": 119456, "author_name": "Özgün Genç", "author_url": "", "post_date": "2016-05-10T12:25:46.473000", "content": "I also like the spy theory. But according to challenge description the objective is to detect subtle day to day changes in satellite images. So it's not about detecting time but about detecting subtle changes and patterns. It looks more like an open ended research problem rather a practical one. The output of the challenge will be novel tools and techniques to analyze such aerial images, and maybe some patterns we discover analyzing these images. I agree that this would have military/intelligence applications.
", "votes": 0, "replies": [] }, { "id": 119410, "author_name": "DarkestKnight", "author_url": "", "post_date": "2016-05-10T02:02:45.087000", "content": "Can someone point to the real objective behind this exercise (Just for motivation)\n+1 for spy theory though!\nThanks!
", "votes": 0, "replies": [] }, { "id": 119248, "author_name": "==>(AL)<==", "author_url": "", "post_date": "2016-05-08T12:17:54.607000", "content": "Have the images been post-processed in any way after acquisition (besides an obvious removal of the metadata)? Rotation, crop, contrast adjustments, anything like that?
", "votes": 0, "replies": [] }, { "id": 119209, "author_name": "HawkWang", "author_url": "", "post_date": "2016-05-08T03:43:50.077000", "content": "[quote=zero zero;118531]
\n\nI agree that this feels like a strange competition (for all reasons listed by the rest of you). For those lamenting not having a titanX, I would not go out and buy one for this competition...given that it's a hard task for a human it would probably be a difficult CNN problem unless you come up with some clever pipeline that somehow isolates important areas...not saying it's impossible but cleverness may win out.
\n\nEveryone has their theories on what is being asked here...
\n\nIt is NOT to improve Draper's space craft landing systems!
\n\nMy pet theory is that it is an intelligence(spy) exercise. What "they" have is many sources of images but they are actually pictures taken of images or computer screens, etc.. The agent(s)(human or otherwise) taking the pictures/video/photocopies/screen-grabs cannot say much about the time stamps, etc.. So "they" are left with a ton of images and a need to time sequence them.
\n\nHa! Like I said, we all have an opinion.
\n\n[/quote]
\n\nI like this spy story!
", "votes": 0, "replies": [] }, { "id": 118134, "author_name": "blindjesse", "author_url": "", "post_date": "2016-05-02T22:11:48.020000", "content": "Just a thought - I wonder if the data is meant to be similar to some sort of drone image set. It seems closer to that than any usual kind of satellite imagery or rigorously-taken aerial photography.
", "votes": 0, "replies": [] }, { "id": 117756, "author_name": "", "author_url": "", "post_date": "2016-04-30T17:58:24.030000", "content": "", "votes": 0, "replies": [] }, { "id": 117603, "author_name": "", "author_url": "", "post_date": "2016-04-29T19:59:46.347000", "content": "", "votes": 0, "replies": [] }, { "id": 117522, "author_name": "", "author_url": "", "post_date": "2016-04-29T14:13:30.373000", "content": "", "votes": 0, "replies": [] }, { "id": 117693, "author_name": "", "author_url": "", "post_date": "2016-04-30T09:17:25.593000", "content": "", "votes": 1, "replies": [] }, { "id": 117672, "author_name": "", "author_url": "", "post_date": "2016-04-30T05:21:17.553000", "content": "", "votes": 9, "replies": [] } ], "raw_markdown_by_id": { "117537": "[quote=SecondPlan;117522]\r\n\r\nI am surprised that Kaggle hosts a lot more image related competitions than before. \r\n\r\n[/quote]\r\n\r\n* Deep learning/convolutional nets are trending. \r\n\r\n* Previous Kaggle image contests saw the state-of-the-art being beaten, or have their solutions adapted for production. \r\n\r\n* Structured data competitions become more-and-more ensembling wars.", "117585": "[quote=inversion;117580]\r\n\r\nI'm assuming this means we can hand label the **test** images.\r\n\r\n[/quote]\r\n\r\nCorrect. We think the best approach is going to be some combination of man + machine. Make sure to document what you do either way!", "117540": "Just for those who missed it when 'reading' the rules, hand labelling **is** allowed in this competiton:\r\n\r\n> Given the unique nature of this competition design, please note a\r\n> couple key changes to Kaggle's standard rules, listed below: \r\n> \r\n> Hand labeling and human prediction are permitted. Winners must deliver\r\n> documentation that clearly outlines the manual and automatic steps\r\n> required to create the winning submission. Use of external data is\r\n> permitted.\r\n\r\nThis should make it rather interesting...", "117529": "[quote=SecondPlan;117522]\r\n\r\nI am surprised that Kaggle hosts a lot more image related competitions than before. \r\n\r\n[/quote]\r\n\r\nYeah, me too, which is annoying because image competitions are not my strong suit. I also have a feeling this will be a very difficult competition, as I can't see anything in the images that would suggest when it was taken.\r\n\r\nIt also seems a little weird that the organisers are posting this as a satellite image chronology competition - one would think that satellites have some sort of clock in them to tell when the photo was taken...", "118531": "I agree that this feels like a strange competition (for all reasons listed by the rest of you). For those lamenting not having a titanX, I would not go out and buy one for this competition...given that it's a hard task for a human it would probably be a difficult CNN problem unless you come up with some clever pipeline that somehow isolates important areas...not saying it's impossible but cleverness may win out.\r\n\r\nEveryone has their theories on what is being asked here...\r\n\r\nIt is NOT to improve Draper's space craft landing systems! \r\n\r\nMy pet theory is that it is an intelligence(spy) exercise. What \"they\" have is many sources of images but they are actually pictures taken of images or computer screens, etc.. The agent(s)(human or otherwise) taking the pictures/video/photocopies/screen-grabs cannot say much about the time stamps, etc.. So \"they\" are left with a ton of images and a need to time sequence them. \r\n\r\nHa! Like I said, we all have an opinion.", "117876": "[quote===>(AL)<==;117674]\r\n\r\nThis all is just feels weird. There is something going on that we are not told ;)\r\n\r\n[/quote]\r\n\r\nI don't see how is this important. We can keep guessing an organizers motivation, for example they may regularly get access to an image series from some third party source, they can not access timestamp data for these images, and they want to be able to put them in series. Or it may be something completely different, but the secrecy doesn't affect very clearly defined task.", "117637": "[quote===>(AL)<==;117635]\r\n\r\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)\r\n\r\n[/quote]\r\n\r\nI wouldn't call randomly guessing the order a productive idea ;)", "117635": "[quote=SecondPlan;117611]\r\n\r\n[quote=Triskelion;117537]\r\n\r\n[quote=SecondPlan;117522]\r\n\r\nI am surprised that Kaggle hosts a lot more image related competitions than before. \r\n\r\n[/quote]\r\n\r\n* Deep learning/convolutional nets are trending. \r\n\r\n* Previous Kaggle image contests saw the state-of-the-art being beaten, or have their solutions adapted for production. \r\n\r\n* Structured data competitions become more-and-more ensembling wars.\r\n\r\n[/quote]\r\n\r\nSo Kaggle 2.0 is coming, where ideas are favored than crazing ensemble. But to do CNN on this task, I think you need a good GPU\r\n\r\n[/quote]\r\n\r\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)", "117601": "[quote=inversion;117591]\r\n\r\nIf I could poke at the rules a bit . . .\r\n\r\nIn previous contests, it really didn't matter if someone cropped train images, using say, Photoshop, because the test images were never touched.\r\n\r\nBut, what if Photoshop is part of the workflow for test images during this contest? Is that a no-go? Or does one need to do the work in GIMP. What if the functionality is not unique to Photoshop? For example, cropping is pretty standard functionality in all image processing software. Will someone be prize-eligible if they cropped in GIMP, but not if they cropped in Photoshop?\r\n\r\nI supposed there's no way to tell, so maybe it's moot. Unless, of course, someone uses Photoshop functionality that is unavailable in open-source image processing software.\r\n\r\n[/quote]\r\n\r\nWe interpret the rules to favor ideas over technicalities. Cropping is so basic an idea it's like like addition or subtraction; we don't care whether you do it in python, your TI-83, or an abacus. The thing you need to avoid is using closed-source, proprietary magic.", "117580": "_Hand labeling and human prediction are permitted._\r\n\r\nHand labeling has always been permitted, on the **training** data. I'm assuming this means we can hand label the **test** images.", "117535": "[quote=anokas;117529]\r\n\r\nIt also seems a little weird that the organisers are posting this as a satellite image chronology competition - one would think that satellites have some sort of clock in them to tell when the photo was taken...\r\n\r\n[/quote]\r\n\r\nhttp://www.draper.com/news/cracking-code-satellite-data\r\n\r\n> Assembling a week’s worth of pictures of the same location in proper chronological order requires contestants to **create and demonstrate new tools** that can be applied to the expected hundreds of terabytes of daily data from the new constellations.\r\n\r\nAlso, perhaps they have a baseline of internal intelligence analysts, and they want to see if crowd-sourcing can beat this.", "117520": "Welcome to Draper's Satellite Image Chronology competition! This one is a little different than our standard fare here on Kaggle. We expect this to be a challenging task for machine and man alike. As such, we are running the competition under rules that reward the best combination of analyst(s) + machine.\r\n\r\nGood luck!", "117674": "[quote=Ben Kamphaus;117672]\r\n\r\nAs someone who worked in the satellite imaging industry for several years I find this to be an odd competition. I'm all about exploring data in new ways or crowd sourcing problems to see if you can find new techniques, but this competition's wording seems to imply that:\r\n\r\n* this is the kind of imagery earth observing satellites typically generate for analysts to use (it isn't)\r\n* there hasn't been much innovation regarding time series recently (there's been a ton, just google scholar search for \"satellite time series imagery\")\r\n* that change detection is hard (in the vast majority of cases it's trivial)\r\n* that this is the kind of task people people need to figure out how to do (it's not, satellite ephemeris data with precise time stamps is pretty much universal).\r\n\r\nSo what's going on here? It's an interesting puzzle, I guess, but most changing time information are at periods that aren't going to necessarily be visible across the dataset:\r\n\r\n* vegetation changes fairly minute over day to day periods, but significant over seasons.\r\n* activity around buildings, traffic, etc. changes for time of day and specific day\r\n* the sensor may move or have moved in a particular path that makes sense?\r\n\r\nBut it's not like this contains the additional band(s) (e.g. near infrared, etc.) you'd want to estimate vegetation cycles with any accuracy. So maybe you could experiment with several images in the test set and do better than random and ... given that hand selection with additional research is ok, how exactly does this result in something that generalizes or pushes the state of the art forward?\r\n\r\nAnyways, I'm gonna do a bit more exploratory analysis and think it over, but this is just a weird feeling contest. Like I'm wondering if there are messages encrypted in the imagery in small noise values that don't impact how the images appear viewed under typical image histogram stretch sensitivities level, and the contest is really about finding and decoding those.\r\n\r\n[/quote]\r\n\r\nMy thoughts exactly. I also spend quite a bit of time with satellite imagery and even thought Satellite Remote Sensing class to master students. The topic says \"satellite imagery\" but it has nothing to do with it in reality. It gives aerophoto instead of satellite - it's like saying that an automobile and an oax cart is the same thing. The both have wheels! it implies that time-stamping of an image is an important task, but I have never seen a satellite image without a timestamp (and geo-location information, sun elevation, etc, in one way or another)\r\n\r\nI have a feel that the real purpose of this puzzle was not the one we were given. Furthermore, it all can be relatively easily done - images can be co-registered, change detection is at least several decades considered a mature science and can be done easily on co-registered images. Nothing of these activities will really move state of the art forward. Maybe sponsors are after new techniques for object localization and recognition - but why not to say so?\r\n\r\nThere is another weird thing about this contest - the target is not well defined. We are asked to produce the right order for images taken over several days. We can, of course, to find that cars in the residential neighborhood are parked differently, for example. I could have imagined that this could have been about time series, and establishing some novel ways of combining RNN with CNN for example. But this is apparently not the case either. We could have used time effects to order things if images were overlapping with some movement, like cars moving on a highway. Or if images were taken over longer time interval when vegetation changes. But images are taken several days apart, in unknown part of the day, so no real time series either. Train and test images are taken from the different areas in the most part. There is very little overlap between train and test. Hand labeling is allowed. \r\n\r\nThis all is just feels weird. There is something going on that we are not told ;)", "118047": "[quote===>(AL)<==;117648]\r\n\r\n[quote=anokas;117637]\r\n\r\n[quote===>(AL)<==;117635]\r\n\r\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)\r\n\r\n[/quote]\r\n\r\nI wouldn't call randomly guessing the order a productive idea ;)\r\n\r\n[/quote]\r\n\r\nIf we define productive as giving people high LB standing then it was a productive idea, at least at the time\r\n\r\n\r\n[/quote]\r\n\r\nInteresting insight about the motives. Maybe someone wants to see if you guys can detect timestamp tampering or not ? ", "117772": "Yes, the metric will give -1 for perfect anticorrelation.", "117648": "[quote=anokas;117637]\r\n\r\n[quote===>(AL)<==;117635]\r\n\r\nMaybe you do not need a GPU. You see, a simple but productive idea already propelled people to the top of the LB ;)\r\n\r\n[/quote]\r\n\r\nI wouldn't call randomly guessing the order a productive idea ;)\r\n\r\n[/quote]\r\n\r\nIf we define productive as giving people high LB standing then it was a productive idea, at least at the time\r\n", "117634": "[quote=William Cukierski;117603]\r\n\r\n[quote=Jmh1;117593]\r\n\r\nJust to clarify, each picture of a set was taken on a different day. They are not the same day, but at different times or angles? \r\n\r\n[/quote]\r\n\r\nCorrect\r\n\r\n[/quote]\r\n\r\nThis wasn't clear from the description that each image is taken on a **different** day? So there is no rapid-shot sequences within few minutes/seconds from the same day? **All** images from a given set are taken at **different** days and possibly from different time of the day, correct? Are they within a short time frame from each other - like days, or some may be month or years apart?", "117611": "[quote=Triskelion;117537]\r\n\r\n[quote=SecondPlan;117522]\r\n\r\nI am surprised that Kaggle hosts a lot more image related competitions than before. \r\n\r\n[/quote]\r\n\r\n* Deep learning/convolutional nets are trending. \r\n\r\n* Previous Kaggle image contests saw the state-of-the-art being beaten, or have their solutions adapted for production. \r\n\r\n* Structured data competitions become more-and-more ensembling wars.\r\n\r\n[/quote]\r\n\r\nSo Kaggle 2.0 is coming, where ideas are favored than crazing ensemble. But to do CNN on this task, I think you need a good GPU", "117593": "Just to clarify, each picture of a set was taken on a different day? They are not the same day, but at different times or angles? ", "117591": "If I could poke at the rules a bit . . .\r\n\r\nIn previous contests, it really didn't matter if someone cropped train images, using say, Photoshop, because the test images were never touched.\r\n\r\nBut, what if Photoshop is part of the workflow for test images during this contest? Is that a no-go? Or does one need to do the work in GIMP. What if the functionality is not unique to Photoshop? For example, cropping is pretty standard functionality in all image processing software. Will someone be prize-eligible if they cropped in GIMP, but not if they cropped in Photoshop?\r\n\r\nI supposed there's no way to tell, so maybe it's moot. Unless, of course, someone uses Photoshop functionality that is unavailable in open-source image processing software.", "117553": "This competition reminds me of the skynet in the movie of Terminator. Fellow Kaggler, are you ready to save the human race?\r\n\r\n\r\nThe trick of this competition is identify the mobile feature, such as car, from fixed structure.\r\n\r\nThe shadow of the building is very interesting. Are the pictures taken exactly the same time during the five day period? ", "117560": "[quote=Birchwood;117553]\r\n\r\nThe shadow of the building is very interesting. Are the pictures taken exactly the same time during the five day period? \r\n\r\n[/quote]\r\n\r\nFrom the data page:\r\n\r\nEach image in a set was taken on a different day (but not necessarily at the same time each day).", "117524": "I convinced that I must buy a titan x.", "117766": "I have a question about metric used for leaderboard. If correct answer is 1 2 3 4 5, but I output 5 4 3 2 1. Will it give -1 as score?\r\n\r\n print(spearmanr([1,2,3,4,5], [1,2,3,4,5]))\r\n print(spearmanr([1,2,3,4,5], [1,2,4,3,5]))\r\n print(spearmanr([1,2,3,4,5], [5,1,2,3,4]))\r\n print(spearmanr([1,2,3,4,5], [5,4,3,2,1]))\r\n print(spearmanr([1,2,3,4,5], [5,4,2,3,1]))\r\n\r\n (1.0, 0.0)\r\n (0.9, 0.037)\r\n (0.0, 1.0)\r\n (-1.0, 0.0)\r\n (-0.9, 0.037)\r\n\r\nI think knowing the right order is actually better than random guess producing the better score.\r\n", "117576": "If I understand this correctly, all the images within one set were taken on the same day ?\r\n", "119456": "I also like the spy theory. But according to challenge description the objective is to detect subtle day to day changes in satellite images. So it's not about detecting time but about detecting subtle changes and patterns. It looks more like an open ended research problem rather a practical one. The output of the challenge will be novel tools and techniques to analyze such aerial images, and maybe some patterns we discover analyzing these images. I agree that this would have military/intelligence applications.", "119410": "Can someone point to the real objective behind this exercise (Just for motivation)\r\n+1 for spy theory though!\r\nThanks!", "119248": "Have the images been post-processed in any way after acquisition (besides an obvious removal of the metadata)? Rotation, crop, contrast adjustments, anything like that?", "119209": "[quote=zero zero;118531]\r\n\r\nI agree that this feels like a strange competition (for all reasons listed by the rest of you). For those lamenting not having a titanX, I would not go out and buy one for this competition...given that it's a hard task for a human it would probably be a difficult CNN problem unless you come up with some clever pipeline that somehow isolates important areas...not saying it's impossible but cleverness may win out.\r\n\r\nEveryone has their theories on what is being asked here...\r\n\r\nIt is NOT to improve Draper's space craft landing systems! \r\n\r\nMy pet theory is that it is an intelligence(spy) exercise. What \"they\" have is many sources of images but they are actually pictures taken of images or computer screens, etc.. The agent(s)(human or otherwise) taking the pictures/video/photocopies/screen-grabs cannot say much about the time stamps, etc.. So \"they\" are left with a ton of images and a need to time sequence them. \r\n\r\nHa! Like I said, we all have an opinion.\r\n\r\n[/quote]\r\n\r\nI like this spy story!", "118134": "Just a thought - I wonder if the data is meant to be similar to some sort of drone image set. It seems closer to that than any usual kind of satellite imagery or rigorously-taken aerial photography.", "117756": "[quote=I can do the splits NO PROB;117693]\r\n\r\n@AL: The untold story? Have you visited draper.com? ;)\r\n\r\n[/quote]\r\n\r\nI have. I've known them before this ;)", "117603": "[quote=Jmh1;117593]\r\n\r\nJust to clarify, each picture of a set was taken on a different day? They are not the same day, but at different times or angles? \r\n\r\n[/quote]\r\n\r\nCorrect", "117522": "I am surprised that Kaggle hosts a lot more image related competitions than before. ", "117693": "", "117672": "" } }