![\"A](\"https://upload.wikimedia.org/wikipedia/commons/0/0e/U2_Image_of_Cuban_Missile_Crisis.jpg\")
I mean, as a human being myself, I have difficulty figuring out which one is taken before another one!
\n\nAny thoughts?
", "messages": [ { "id": "117558", "postDate": "04/29/2016 16:36:48", "content": "I mean, as a human being myself, I have difficulty figuring out which one is taken before another one!
\n\nAny thoughts?
", "rawMarkdown": "I mean, as a human being myself, I have difficulty figuring out which one is taken before another one!\r\n\r\nAny thoughts?", "votes": null }, { "id": "117571", "postDate": "04/29/2016 17:44:49", "content": "I think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point?
", "rawMarkdown": "I think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point?", "votes": null }, { "id": "117574", "postDate": "04/29/2016 18:01:01", "content": "[quote=Gerard Toonstra;117571]
\n\nI think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point?
\n\n[/quote]
\n\nI think this competition is not actually for satellite imagery, but to try and find novel algorithms for image analysis.
", "rawMarkdown": "[quote=Gerard Toonstra;117571]\r\n\r\nI think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point? \r\n\r\n[/quote]\r\n\r\nI think this competition is not actually for satellite imagery, but to try and find novel algorithms for image analysis.", "votes": null }, { "id": "117708", "postDate": "04/30/2016 13:25:35", "content": "[quote=Gerard Toonstra;117571]\nI think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point? \n[/quote]
\n\nIf the organizers of this competition were smart (reasonable preposition), they would make this kind of info irrelevant to the image ordering, together with the shadow size, image zoom and rotation factor, etc.
\n\nSeems to me that the spirit of the competition would force the ability to figure out what is happening with the objects in the image the key to successfully order the images. Everything else would be essentially a data leak.
", "rawMarkdown": "[quote=Gerard Toonstra;117571]\r\nI think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point? \r\n[/quote]\r\n\r\nIf the organizers of this competition were smart (reasonable preposition), they would make this kind of info irrelevant to the image ordering, together with the shadow size, image zoom and rotation factor, etc.\r\n\r\nSeems to me that the spirit of the competition would force the ability to figure out what is happening with the objects in the image the key to successfully order the images. Everything else would be essentially a data leak.", "votes": null }, { "id": "117776", "postDate": "04/30/2016 21:05:41", "content": "[quote=Gerard Toonstra;117571]
\n\nI think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point?
\n\n[/quote]
\n\nThis information is not always available. For example, if you hoover up images off the internet.
\n\nInstagram, Facebook, Twitter, and Skype routinely rename image files and remove headers, metadata, and all context (EXIF, IPTC, etc). (Their reasons are not to protect your privacy, but to keep this useful information to themselves. Google Photos does not do this.)
\n\nI do not think that the "satellite" part of this competition is important. It's mostly the image chronology that Draper are interested in.
", "rawMarkdown": "[quote=Gerard Toonstra;117571]\r\n\r\nI think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point? \r\n\r\n[/quote]\r\n\r\nThis information is not always available. For example, if you hoover up images off the internet. \r\n\r\nInstagram, Facebook, Twitter, and Skype routinely rename image files and remove headers, metadata, and **all context** (EXIF, IPTC, etc). (Their reasons are not to protect your privacy, but to keep this useful information to themselves. Google Photos does not do this.)\r\n\r\nI do not think that the \"satellite\" part of this competition is important. It's mostly the image chronology that Draper are interested in.", "votes": null }, { "id": "117782", "postDate": "04/30/2016 21:34:34", "content": "Except that these images aren't downloaded from the internet and usually would have a lot of metadata embedded in them that's kept there. I was the business owner of a company that used consumer drones to run topography missions for very small areas. The cameras we used were Canon portable handheld ones that your auntie would buy at BestBuy, S100, S110, S260. Even those cameras keep internal time that keeps the time to a second precise. On many cameras with GPS, you don't even need to synchronize, because it syncs to GPS time. So I don't see how Draper wouldn't have access to such image metadata and is really stuck with the question of image chronology.
\n\nI reckon the competition is actually about figuring out movements and changes instead, but formulating that as a labeling question is probably very hard. The idea is that algorithms capable of predicting chronology should be useful to threshold on image changes as well and rewire them to flag anomalies.
", "rawMarkdown": "Except that these images aren't downloaded from the internet and usually would have a lot of metadata embedded in them that's kept there. I was the business owner of a company that used consumer drones to run topography missions for very small areas. The cameras we used were Canon portable handheld ones that your auntie would buy at BestBuy, S100, S110, S260. Even those cameras keep internal time that keeps the time to a second precise. On many cameras with GPS, you don't even need to synchronize, because it syncs to GPS time. So I don't see how Draper wouldn't have access to such image metadata and is really stuck with the question of image chronology.\r\n\r\nI reckon the competition is actually about figuring out movements and changes instead, but formulating that as a labeling question is probably very hard. The idea is that algorithms capable of predicting chronology should be useful to threshold on image changes as well and rewire them to flag anomalies.", "votes": null }, { "id": "117799", "postDate": "05/01/2016 00:38:27", "content": "[quote=Gerard Toonstra;117782]\nI reckon the competition is actually about figuring out movements and changes instead, but formulating that as a labeling question is probably very hard. The idea is that algorithms capable of predicting chronology should be useful to threshold on image changes as well and rewire them to flag anomalies.\n[/quote]
\n\nExactly!
\n\nRemember this?
\n\n
Edit: https://en.wikipedia.org/wiki/Cuban_Missile_Crisis
", "rawMarkdown": "[quote=Gerard Toonstra;117782]\r\nI reckon the competition is actually about figuring out movements and changes instead, but formulating that as a labeling question is probably very hard. The idea is that algorithms capable of predicting chronology should be useful to threshold on image changes as well and rewire them to flag anomalies.\r\n[/quote]\r\n\r\nExactly!\r\n\r\nRemember this?\r\n\r\n![A U-2 reconnaissance photograph of Cuba, showing Soviet nuclear missiles, their transports and tents for fueling and maintenance.][1]\r\n\r\nEdit: https://en.wikipedia.org/wiki/Cuban_Missile_Crisis\r\n\r\n\r\n [1]: https://upload.wikimedia.org/wikipedia/commons/0/0e/U2_Image_of_Cuban_Missile_Crisis.jpg", "votes": null }, { "id": "117854", "postDate": "05/01/2016 08:16:03", "content": "Anyone have ideas how to detect which image was first and which one was last? Five Days is too small period to have some noticible changes in trees, buildings etc.
", "rawMarkdown": "Anyone have ideas how to detect which image was first and which one was last? Five Days is too small period to have some noticible changes in trees, buildings etc.", "votes": null }, { "id": "117857", "postDate": "05/01/2016 08:29:23", "content": "Well, there may be other ways about this, but my initial intent is to coregister and reproject them. Because there's a slight oblique angle here and there, even if you figure out how to translate them into the same image space, it's likely that not all the pixels match up.
\n\nYesterday I tried, for two sets, to draw all the images into the same space, but this didn't work out. I didn't have a lot of time, but it appears that the images that are closer to earth are not necessarily taken at lower altitudes, but may have different zoom levels. This can have an impact, because the camera parameters may change significantly between them.
\n\nI only tried two sets, but it's already a word of warning that, irrespective of the immense detail, getting them to co-register properly may already be a daunting task.
", "rawMarkdown": "Well, there may be other ways about this, but my initial intent is to coregister and reproject them. Because there's a slight oblique angle here and there, even if you figure out how to translate them into the same image space, it's likely that not all the pixels match up. \r\n\r\nYesterday I tried, for two sets, to draw all the images into the same space, but this didn't work out. I didn't have a lot of time, but it appears that the images that are closer to earth are not necessarily taken at lower altitudes, but may have different zoom levels. This can have an impact, because the camera parameters may change significantly between them.\r\n\r\nI only tried two sets, but it's already a word of warning that, irrespective of the immense detail, getting them to co-register properly may already be a daunting task.", "votes": null }, { "id": "117968", "postDate": "05/02/2016 04:38:47", "content": "My current public LB score (0.29) is based on the pure ML approach (no hand labelling, no external data, no random guessing, no meta data...). I extracted features from each train set and applied to test set by Random Forest.
\n\nThe strange thing is that my local CV score is not so bad, but it miserably failed on the test set. I might be doing something wrong, but it feels like there fundamental difference between them. (Is this the reason why the admin said that winning solution must be half man half machine approach?)
", "rawMarkdown": "My current public LB score (0.29) is based on the pure ML approach (no hand labelling, no external data, no random guessing, no meta data...). I extracted features from each train set and applied to test set by Random Forest.\r\n\r\nThe strange thing is that my local CV score is not so bad, but it miserably failed on the test set. I might be doing something wrong, but it feels like there fundamental difference between them. (Is this the reason why the admin said that winning solution must be half man half machine approach?)", "votes": null }, { "id": "118029", "postDate": "05/02/2016 16:20:13", "content": "[quote=ZFTurbo;117854]
\n\nAnyone have ideas how to detect which image was first and which one was last? Five Days is too small period to have some noticible changes in trees, buildings etc.
\n\n[/quote]
\n\nUse satellite images to align all photos to true north, and to get their latitude. Use the shadow angle to calculate the time of day. Use the time of day, latitude, and shadow length to calculate solar elevation angle at noon. Sort by decreasing or increasing elevation, depending on whether the photos were taken in spring or in fall.
", "rawMarkdown": "[quote=ZFTurbo;117854]\r\n\r\nAnyone have ideas how to detect which image was first and which one was last? Five Days is too small period to have some noticible changes in trees, buildings etc.\r\n\r\n[/quote]\r\n\r\nUse satellite images to align all photos to true north, and to get their latitude. Use the shadow angle to calculate the time of day. Use the time of day, latitude, and shadow length to calculate solar elevation angle at noon. Sort by decreasing or increasing elevation, depending on whether the photos were taken in spring or in fall.", "votes": null }, { "id": "118106", "postDate": "05/02/2016 21:36:07", "content": "Any idea how to find images on Google Earth except manual labeling?
", "rawMarkdown": "Any idea how to find images on Google Earth except manual labeling?", "votes": null }, { "id": "119018", "postDate": "05/06/2016 18:26:04", "content": "\n\n\nUse satellite images to align all photos to true north, and to get their latitude. Use the shadow angle to calculate the time of day. Use the time of day, latitude, and shadow length to calculate solar elevation angle at noon. Sort by decreasing or increasing elevation, depending on whether the photos were taken in spring or in fall.
\n
Is that kind of approach allowed in this competition? (Assuming it was feasible)
", "rawMarkdown": "> Use satellite images to align all photos to true north, and to get their latitude. Use the shadow angle to calculate the time of day. Use the time of day, latitude, and shadow length to calculate solar elevation angle at noon. Sort by decreasing or increasing elevation, depending on whether the photos were taken in spring or in fall.\r\n\r\nIs that kind of approach allowed in this competition? (Assuming it was feasible)", "votes": null }, { "id": "119254", "postDate": "05/08/2016 14:58:42", "content": "[quote=MattZhang;119018]
\n\n\n\n\nUse satellite images to align all photos to true north, and to get their latitude. Use the shadow angle to calculate the time of day. Use the time of day, latitude, and shadow length to calculate solar elevation angle at noon. Sort by decreasing or increasing elevation, depending on whether the photos were taken in spring or in fall.
\n
Is that kind of approach allowed in this competition? (Assuming it was feasible)
\n\n[/quote]
\n\nI think this is a man+machine approach and is should be allowed. Do we need the true north and the exact location though? All the images are from the same region and taken at the same days. And we have a training set with true orderings. I don't know but maybe there is enough information in this setup.
", "rawMarkdown": "[quote=MattZhang;119018]\r\n\r\n> Use satellite images to align all photos to true north, and to get their latitude. Use the shadow angle to calculate the time of day. Use the time of day, latitude, and shadow length to calculate solar elevation angle at noon. Sort by decreasing or increasing elevation, depending on whether the photos were taken in spring or in fall.\r\n\r\nIs that kind of approach allowed in this competition? (Assuming it was feasible)\r\n\r\n[/quote]\r\n\r\nI think this is a man+machine approach and is should be allowed. Do we need the true north and the exact location though? All the images are from the same region and taken at the same days. And we have a training set with true orderings. I don't know but maybe there is enough information in this setup.", "votes": null }, { "id": "119260", "postDate": "05/08/2016 16:23:38", "content": "| Do we need the true north and the exact location though?
\n\nMy suggestion was only half-serious. To check the feasibility of the approach, I just used an online solar elevation calculator to calculate the elevation at noon today and tomorrow in San Diego. I got 74.25 and 74.51 degrees. If I did my math correctly, this means the length of a shadow of a 10 meter long vertical object changes by 0.05 meters per day.
", "rawMarkdown": "| Do we need the true north and the exact location though?\r\n\r\nMy suggestion was only half-serious. To check the feasibility of the approach, I just used an online solar elevation calculator to calculate the elevation at noon today and tomorrow in San Diego. I got 74.25 and 74.51 degrees. If I did my math correctly, this means the length of a shadow of a 10 meter long vertical object changes by 0.05 meters per day.", "votes": null }, { "id": "119435", "postDate": "05/10/2016 08:41:36", "content": "I think it's a interesting one. It'll need a novel idea to make any real progress I think.
", "rawMarkdown": "I think it's a interesting one. It'll need a novel idea to make any real progress I think.", "votes": null }, { "id": "119898", "postDate": "05/13/2016 16:10:10", "content": "[quote=khyh;117968]
\n\nThe strange thing is that my local CV score is not so bad, but it miserably failed on the test set. I might be doing something wrong, but it feels like there fundamental difference between them. (Is this the reason why the admin said that winning solution must be half man half machine approach?)
\n\n[/quote]
\n\nI had a local 0.55 CV yet it got crushed to 0.07 in public LB (I did not even look at the picture sets - and I had tied ranks I did not fix too, which complicated the issue of interpreting it due to Spearman rank correlation as it is clearly not explained how ties are handled). Looking at the images directly it seems there are a lot of different types of pictures to predict on. We do not have many themes given in the train set, which is quite a pain (building? water? where are forests? nature? some types of "strange" building, erm?). It might be possible it could predict a bit on the themes given in the train set, but failed really hard on the unknown thematics provided in the test set.
\n\nP.S: I took 5 random sets in the train set and predicted manually at 4 of them perfectly. Hand labeling correction in this competition seem to be potentially much better than ML alone, have to test that on more sets (but it's a pain to order them 1 by 1).
\n\nAlso, it's possible to reduce the problem into a pairwise ranking problem: for each pair in the set, order the pictures by pair. Once you have all the pair rules, order the whole 5 pictures of the set by rule precedence (i.e the stronger rules supersedes the weaker rules in case of rule conflicts).
", "rawMarkdown": "[quote=khyh;117968]\r\n\r\nThe strange thing is that my local CV score is not so bad, but it miserably failed on the test set. I might be doing something wrong, but it feels like there fundamental difference between them. (Is this the reason why the admin said that winning solution must be half man half machine approach?)\r\n\r\n[/quote]\r\n\r\nI had a local 0.55 CV yet it got crushed to 0.07 in public LB (I did not even look at the picture sets - and I had tied ranks I did not fix too, which complicated the issue of interpreting it due to Spearman rank correlation as it is clearly not explained how ties are handled). Looking at the images directly it seems there are a lot of different types of pictures to predict on. We do not have many themes given in the train set, which is quite a pain (building? water? where are forests? nature? some types of \"strange\" building, erm?). It might be possible it could predict a bit on the themes given in the train set, but failed really hard on the unknown thematics provided in the test set.\r\n\r\nP.S: I took 5 random sets in the train set and predicted manually at 4 of them perfectly. Hand labeling correction in this competition seem to be potentially much better than ML alone, have to test that on more sets (but it's a pain to order them 1 by 1).\r\n\r\n* Providing a ML solution: Draper will be interested in, \"pre-cooked\".\r\n* Providing a hand-classified solution: Draper will be really much interested at the framework you created to order the pictures, and they may find a way to translate the problem into a more accurate ML problem to solve by their own means (the framework being usually the preliminary/precursor of a larger research).\r\n\r\nAlso, it's possible to reduce the problem into a pairwise ranking problem: for each pair in the set, order the pictures by pair. Once you have all the pair rules, order the whole 5 pictures of the set by rule precedence (i.e the stronger rules supersedes the weaker rules in case of rule conflicts).", "votes": null }, { "id": "119919", "postDate": "05/13/2016 18:47:52", "content": "@ Laurae, \nMay I ask on what criteria you managed to predict the 4 sets out of 5 ? what are you looking for ? I thought the change in the orientation of the shade and the length of the shade will be the key given the train set but that give me headaches for most of the pictures in the test set.
\n\nMaybe a manual classification into residential, highways, river, industry, harbor, forest could help when aggregated with the dataset such as the one built by the1owl script to feed a ML classifier.
\n\nI finally understood the "registration" show by Ben Kamphaus in his amazing script but I haven't figure how to resize the all set into the overlapping area yet. Do you think we can avoid re-sizing of the overlapping area for each set of picture in this challenge ?
", "rawMarkdown": "Laurae, \r\nMay I ask on what criteria you managed to predict the 4 sets out of 5 ? what are you looking for ? I thought the change in the orientation of the shade and the length of the shade will be the key given the train set but that give me headaches for most of the pictures in the test set.\r\n\r\nMaybe a manual classification into residential, highways, river, industry, harbor, forest could help when aggregated with the dataset such as the one built by the1owl script to feed a ML classifier.\r\n\r\nI finally understood the \"registration\" show by Ben Kamphaus in his amazing script but I haven't figure how to resize the all set into the overlapping area yet. Do you think we can avoid re-sizing of the overlapping area for each set of picture in this challenge ?", "votes": null }, { "id": "119929", "postDate": "05/13/2016 20:11:10", "content": "[quote=eagle4;119919]
\n\n@ Laurae, \nMay I ask on what criteria you managed to predict the 4 sets out of 5 ? what are you looking for ? I thought the change in the orientation of the shade and the length of the shade will be the key given the train set but that give me headaches for most of the pictures in the test set.
\n\n[/quote]
\n\nFirst, I used registration alignment (to align using an affine transformation) then elastic registration (to scale using a proximity transformation) to make sure everything is well aligned. Then, as each images are "closely" aligned on every object (you have many ways to do that, it requires to use a keypoint detector such as KAZE, Harris corner detection, etc.), I focus on the very details of the each image by swapping images very quickly so my eyes can catch up potential strange differences (remember "strange": for instance if cars disappear in the middle of a street, it's perfectly normal). Once I spotted an area with major detail changes, I look for the same area for the five pictures.
\n\nAssuming we have only two pictures (A and B) with that detail change, we can conclude either one or more of the following:
\n\nAnd I do this until I have enough rules to conclude about the complete order of the pictures.
\n\nI tried to use the shade length/orientation, I found it can swap two or three pictures (due to unperfect alignment/scaling).
\n\nI had issues like the following when reproducing the theoretical Z-axis (scale = order of the picture), example:
\n\n![\"enter](\"http://i.imgur.com/80Oc8DP.png\")
Using only the shade, I would have predicted 4, 1, 5, 2, 3, which is wrong (and gives -0.1 Spearman rank correlation).
\n\nThe major issue is finding the right details to find the right order of pictures, because sometimes it's either:
\n\nAlso, I am assuming the pictures are not taken at the exact time of day each (is it the case?). And one cloudy/windy weather would be enough to dampen the usage of shades (there are pictures with severe clouds or wind).
\n\nBy the way, how do you deal with 3099x2329 pictures? Do you crop 1 horizontal pixel from 3100x2329 pictures?
", "rawMarkdown": "[quote=eagle4;119919]\r\n\r\n@ Laurae, \r\nMay I ask on what criteria you managed to predict the 4 sets out of 5 ? what are you looking for ? I thought the change in the orientation of the shade and the length of the shade will be the key given the train set but that give me headaches for most of the pictures in the test set.\r\n\r\n[/quote]\r\n\r\nFirst, I used registration alignment (to align using an affine transformation) then elastic registration (to scale using a proximity transformation) to make sure everything is well aligned. Then, as each images are \"closely\" aligned on every object (you have many ways to do that, it requires to use a keypoint detector such as KAZE, Harris corner detection, etc.), I focus on the very details of the each image by swapping images very quickly so my eyes can catch up potential strange differences (remember \"strange\": for instance if cars disappear in the middle of a street, it's perfectly normal). Once I spotted an area with major detail changes, I look for the same area for the five pictures.\r\n\r\nAssuming we have only two pictures (A and B) with that detail change, we can conclude either one or more of the following:\r\n\r\n* A > B\r\n* B < A\r\n* A > B or B < A\r\n* A = 1 or B = 1\r\n* A = 5 or B = 5\r\n* A = 1 or B = 1 or A = 5 or B = 5\r\n\r\nAnd I do this until I have enough rules to conclude about the complete order of the pictures.\r\n\r\nI tried to use the shade length/orientation, I found it can swap two or three pictures (due to unperfect alignment/scaling).\r\n\r\nI had issues like the following when reproducing the theoretical Z-axis (scale = order of the picture), example:\r\n\r\n![enter image description here][1]\r\n\r\nUsing only the shade, I would have predicted 4, 1, 5, 2, 3, which is wrong (and gives -0.1 Spearman rank correlation).\r\n\r\nThe major issue is finding the right details to find the right order of pictures, because sometimes it's either:\r\n\r\n* A, B, C, D, E\r\n* E, D, C, B, A\r\n\r\nAlso, I am assuming the pictures are not taken at the exact time of day each (is it the case?). And one cloudy/windy weather would be enough to dampen the usage of shades (there are pictures with severe clouds or wind).\r\n\r\nBy the way, how do you deal with 3099x2329 pictures? Do you crop 1 horizontal pixel from 3100x2329 pictures?\r\n\r\n [1]: http://i.imgur.com/80Oc8DP.png", "votes": null }, { "id": "119998", "postDate": "05/14/2016 14:00:09", "content": "Re ABCDE vs EDCBA: this is especially annoying given the scoring algorithm, since EDCBA results in -1.0 even though logically you have the correct sequence.
", "rawMarkdown": "Re ABCDE vs EDCBA: this is especially annoying given the scoring algorithm, since EDCBA results in -1.0 even though logically you have the correct sequence.", "votes": null } ], "comments": [ { "id": 117571, "author_name": "remap1", "author_url": "", "post_date": "04/29/2016 17:44:49", "content": "I think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point?
", "votes": null, "replies": [] }, { "id": 117574, "author_name": "anokas", "author_url": "", "post_date": "04/29/2016 18:01:01", "content": "[quote=Gerard Toonstra;117571]
\n\nI think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point?
\n\n[/quote]
\n\nI think this competition is not actually for satellite imagery, but to try and find novel algorithms for image analysis.
", "votes": null, "replies": [] }, { "id": 117708, "author_name": "itodorovic", "author_url": "", "post_date": "04/30/2016 13:25:35", "content": "[quote=Gerard Toonstra;117571]\nI think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point? \n[/quote]
\n\nIf the organizers of this competition were smart (reasonable preposition), they would make this kind of info irrelevant to the image ordering, together with the shadow size, image zoom and rotation factor, etc.
\n\nSeems to me that the spirit of the competition would force the ability to figure out what is happening with the objects in the image the key to successfully order the images. Everything else would be essentially a data leak.
", "votes": null, "replies": [] }, { "id": 117776, "author_name": "innerproduct", "author_url": "", "post_date": "04/30/2016 21:05:41", "content": "[quote=Gerard Toonstra;117571]
\n\nI think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point?
\n\n[/quote]
\n\nThis information is not always available. For example, if you hoover up images off the internet.
\n\nInstagram, Facebook, Twitter, and Skype routinely rename image files and remove headers, metadata, and all context (EXIF, IPTC, etc). (Their reasons are not to protect your privacy, but to keep this useful information to themselves. Google Photos does not do this.)
\n\nI do not think that the "satellite" part of this competition is important. It's mostly the image chronology that Draper are interested in.
", "votes": null, "replies": [] }, { "id": 117782, "author_name": "remap1", "author_url": "", "post_date": "04/30/2016 21:34:34", "content": "Except that these images aren't downloaded from the internet and usually would have a lot of metadata embedded in them that's kept there. I was the business owner of a company that used consumer drones to run topography missions for very small areas. The cameras we used were Canon portable handheld ones that your auntie would buy at BestBuy, S100, S110, S260. Even those cameras keep internal time that keeps the time to a second precise. On many cameras with GPS, you don't even need to synchronize, because it syncs to GPS time. So I don't see how Draper wouldn't have access to such image metadata and is really stuck with the question of image chronology.
\n\nI reckon the competition is actually about figuring out movements and changes instead, but formulating that as a labeling question is probably very hard. The idea is that algorithms capable of predicting chronology should be useful to threshold on image changes as well and rewire them to flag anomalies.
", "votes": null, "replies": [] }, { "id": 117799, "author_name": "radustoicescu", "author_url": "", "post_date": "05/01/2016 00:38:27", "content": "[quote=Gerard Toonstra;117782]\nI reckon the competition is actually about figuring out movements and changes instead, but formulating that as a labeling question is probably very hard. The idea is that algorithms capable of predicting chronology should be useful to threshold on image changes as well and rewire them to flag anomalies.\n[/quote]
\n\nExactly!
\n\nRemember this?
\n\n
Edit: https://en.wikipedia.org/wiki/Cuban_Missile_Crisis
", "votes": null, "replies": [] }, { "id": 117854, "author_name": "zfturbo", "author_url": "", "post_date": "05/01/2016 08:16:03", "content": "Anyone have ideas how to detect which image was first and which one was last? Five Days is too small period to have some noticible changes in trees, buildings etc.
", "votes": null, "replies": [] }, { "id": 117857, "author_name": "remap1", "author_url": "", "post_date": "05/01/2016 08:29:23", "content": "Well, there may be other ways about this, but my initial intent is to coregister and reproject them. Because there's a slight oblique angle here and there, even if you figure out how to translate them into the same image space, it's likely that not all the pixels match up.
\n\nYesterday I tried, for two sets, to draw all the images into the same space, but this didn't work out. I didn't have a lot of time, but it appears that the images that are closer to earth are not necessarily taken at lower altitudes, but may have different zoom levels. This can have an impact, because the camera parameters may change significantly between them.
\n\nI only tried two sets, but it's already a word of warning that, irrespective of the immense detail, getting them to co-register properly may already be a daunting task.
", "votes": null, "replies": [] }, { "id": 117968, "author_name": "khyh00", "author_url": "", "post_date": "05/02/2016 04:38:47", "content": "My current public LB score (0.29) is based on the pure ML approach (no hand labelling, no external data, no random guessing, no meta data...). I extracted features from each train set and applied to test set by Random Forest.
\n\nThe strange thing is that my local CV score is not so bad, but it miserably failed on the test set. I might be doing something wrong, but it feels like there fundamental difference between them. (Is this the reason why the admin said that winning solution must be half man half machine approach?)
", "votes": null, "replies": [] }, { "id": 118029, "author_name": "mortehu", "author_url": "", "post_date": "05/02/2016 16:20:13", "content": "[quote=ZFTurbo;117854]
\n\nAnyone have ideas how to detect which image was first and which one was last? Five Days is too small period to have some noticible changes in trees, buildings etc.
\n\n[/quote]
\n\nUse satellite images to align all photos to true north, and to get their latitude. Use the shadow angle to calculate the time of day. Use the time of day, latitude, and shadow length to calculate solar elevation angle at noon. Sort by decreasing or increasing elevation, depending on whether the photos were taken in spring or in fall.
", "votes": null, "replies": [] }, { "id": 118106, "author_name": "itodorovic", "author_url": "", "post_date": "05/02/2016 21:36:07", "content": "Any idea how to find images on Google Earth except manual labeling?
", "votes": null, "replies": [] }, { "id": 119018, "author_name": "bucketoffish", "author_url": "", "post_date": "05/06/2016 18:26:04", "content": "\n\n\nUse satellite images to align all photos to true north, and to get their latitude. Use the shadow angle to calculate the time of day. Use the time of day, latitude, and shadow length to calculate solar elevation angle at noon. Sort by decreasing or increasing elevation, depending on whether the photos were taken in spring or in fall.
\n
Is that kind of approach allowed in this competition? (Assuming it was feasible)
", "votes": null, "replies": [] }, { "id": 119254, "author_name": "ozgung", "author_url": "", "post_date": "05/08/2016 14:58:42", "content": "[quote=MattZhang;119018]
\n\n\n\n\nUse satellite images to align all photos to true north, and to get their latitude. Use the shadow angle to calculate the time of day. Use the time of day, latitude, and shadow length to calculate solar elevation angle at noon. Sort by decreasing or increasing elevation, depending on whether the photos were taken in spring or in fall.
\n
Is that kind of approach allowed in this competition? (Assuming it was feasible)
\n\n[/quote]
\n\nI think this is a man+machine approach and is should be allowed. Do we need the true north and the exact location though? All the images are from the same region and taken at the same days. And we have a training set with true orderings. I don't know but maybe there is enough information in this setup.
", "votes": null, "replies": [] }, { "id": 119260, "author_name": "mortehu", "author_url": "", "post_date": "05/08/2016 16:23:38", "content": "| Do we need the true north and the exact location though?
\n\nMy suggestion was only half-serious. To check the feasibility of the approach, I just used an online solar elevation calculator to calculate the elevation at noon today and tomorrow in San Diego. I got 74.25 and 74.51 degrees. If I did my math correctly, this means the length of a shadow of a 10 meter long vertical object changes by 0.05 meters per day.
", "votes": null, "replies": [] }, { "id": 119435, "author_name": "hairykrishna", "author_url": "", "post_date": "05/10/2016 08:41:36", "content": "I think it's a interesting one. It'll need a novel idea to make any real progress I think.
", "votes": null, "replies": [] }, { "id": 119898, "author_name": "laurae2", "author_url": "", "post_date": "05/13/2016 16:10:10", "content": "[quote=khyh;117968]
\n\nThe strange thing is that my local CV score is not so bad, but it miserably failed on the test set. I might be doing something wrong, but it feels like there fundamental difference between them. (Is this the reason why the admin said that winning solution must be half man half machine approach?)
\n\n[/quote]
\n\nI had a local 0.55 CV yet it got crushed to 0.07 in public LB (I did not even look at the picture sets - and I had tied ranks I did not fix too, which complicated the issue of interpreting it due to Spearman rank correlation as it is clearly not explained how ties are handled). Looking at the images directly it seems there are a lot of different types of pictures to predict on. We do not have many themes given in the train set, which is quite a pain (building? water? where are forests? nature? some types of "strange" building, erm?). It might be possible it could predict a bit on the themes given in the train set, but failed really hard on the unknown thematics provided in the test set.
\n\nP.S: I took 5 random sets in the train set and predicted manually at 4 of them perfectly. Hand labeling correction in this competition seem to be potentially much better than ML alone, have to test that on more sets (but it's a pain to order them 1 by 1).
\n\nAlso, it's possible to reduce the problem into a pairwise ranking problem: for each pair in the set, order the pictures by pair. Once you have all the pair rules, order the whole 5 pictures of the set by rule precedence (i.e the stronger rules supersedes the weaker rules in case of rule conflicts).
", "votes": null, "replies": [] }, { "id": 119919, "author_name": "chabir", "author_url": "", "post_date": "05/13/2016 18:47:52", "content": "@ Laurae, \nMay I ask on what criteria you managed to predict the 4 sets out of 5 ? what are you looking for ? I thought the change in the orientation of the shade and the length of the shade will be the key given the train set but that give me headaches for most of the pictures in the test set.
\n\nMaybe a manual classification into residential, highways, river, industry, harbor, forest could help when aggregated with the dataset such as the one built by the1owl script to feed a ML classifier.
\n\nI finally understood the "registration" show by Ben Kamphaus in his amazing script but I haven't figure how to resize the all set into the overlapping area yet. Do you think we can avoid re-sizing of the overlapping area for each set of picture in this challenge ?
", "votes": null, "replies": [] }, { "id": 119929, "author_name": "laurae2", "author_url": "", "post_date": "05/13/2016 20:11:10", "content": "[quote=eagle4;119919]
\n\n@ Laurae, \nMay I ask on what criteria you managed to predict the 4 sets out of 5 ? what are you looking for ? I thought the change in the orientation of the shade and the length of the shade will be the key given the train set but that give me headaches for most of the pictures in the test set.
\n\n[/quote]
\n\nFirst, I used registration alignment (to align using an affine transformation) then elastic registration (to scale using a proximity transformation) to make sure everything is well aligned. Then, as each images are "closely" aligned on every object (you have many ways to do that, it requires to use a keypoint detector such as KAZE, Harris corner detection, etc.), I focus on the very details of the each image by swapping images very quickly so my eyes can catch up potential strange differences (remember "strange": for instance if cars disappear in the middle of a street, it's perfectly normal). Once I spotted an area with major detail changes, I look for the same area for the five pictures.
\n\nAssuming we have only two pictures (A and B) with that detail change, we can conclude either one or more of the following:
\n\nAnd I do this until I have enough rules to conclude about the complete order of the pictures.
\n\nI tried to use the shade length/orientation, I found it can swap two or three pictures (due to unperfect alignment/scaling).
\n\nI had issues like the following when reproducing the theoretical Z-axis (scale = order of the picture), example:
\n\n
Using only the shade, I would have predicted 4, 1, 5, 2, 3, which is wrong (and gives -0.1 Spearman rank correlation).
\n\nThe major issue is finding the right details to find the right order of pictures, because sometimes it's either:
\n\nAlso, I am assuming the pictures are not taken at the exact time of day each (is it the case?). And one cloudy/windy weather would be enough to dampen the usage of shades (there are pictures with severe clouds or wind).
\n\nBy the way, how do you deal with 3099x2329 pictures? Do you crop 1 horizontal pixel from 3100x2329 pictures?
", "votes": null, "replies": [] }, { "id": 119998, "author_name": "tylervigen", "author_url": "", "post_date": "05/14/2016 14:00:09", "content": "Re ABCDE vs EDCBA: this is especially annoying given the scoring algorithm, since EDCBA results in -1.0 even though logically you have the correct sequence.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "117558": "I mean, as a human being myself, I have difficulty figuring out which one is taken before another one!\r\n\r\nAny thoughts?", "117571": "I think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point?", "117574": "[quote=Gerard Toonstra;117571]\r\n\r\nI think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point? \r\n\r\n[/quote]\r\n\r\nI think this competition is not actually for satellite imagery, but to try and find novel algorithms for image analysis.", "117708": "[quote=Gerard Toonstra;117571]\r\nI think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point? \r\n[/quote]\r\n\r\nIf the organizers of this competition were smart (reasonable preposition), they would make this kind of info irrelevant to the image ordering, together with the shadow size, image zoom and rotation factor, etc.\r\n\r\nSeems to me that the spirit of the competition would force the ability to figure out what is happening with the objects in the image the key to successfully order the images. Everything else would be essentially a data leak.", "117776": "[quote=Gerard Toonstra;117571]\r\n\r\nI think it's a really weird question. Why is it not possible to capture the UTC time when a photo was taken and use that as a starting point? \r\n\r\n[/quote]\r\n\r\nThis information is not always available. For example, if you hoover up images off the internet. \r\n\r\nInstagram, Facebook, Twitter, and Skype routinely rename image files and remove headers, metadata, and **all context** (EXIF, IPTC, etc). (Their reasons are not to protect your privacy, but to keep this useful information to themselves. Google Photos does not do this.)\r\n\r\nI do not think that the \"satellite\" part of this competition is important. It's mostly the image chronology that Draper are interested in.", "117782": "Except that these images aren't downloaded from the internet and usually would have a lot of metadata embedded in them that's kept there. I was the business owner of a company that used consumer drones to run topography missions for very small areas. The cameras we used were Canon portable handheld ones that your auntie would buy at BestBuy, S100, S110, S260. Even those cameras keep internal time that keeps the time to a second precise. On many cameras with GPS, you don't even need to synchronize, because it syncs to GPS time. So I don't see how Draper wouldn't have access to such image metadata and is really stuck with the question of image chronology.\r\n\r\nI reckon the competition is actually about figuring out movements and changes instead, but formulating that as a labeling question is probably very hard. The idea is that algorithms capable of predicting chronology should be useful to threshold on image changes as well and rewire them to flag anomalies.", "117799": "[quote=Gerard Toonstra;117782]\r\nI reckon the competition is actually about figuring out movements and changes instead, but formulating that as a labeling question is probably very hard. The idea is that algorithms capable of predicting chronology should be useful to threshold on image changes as well and rewire them to flag anomalies.\r\n[/quote]\r\n\r\nExactly!\r\n\r\nRemember this?\r\n\r\n![A U-2 reconnaissance photograph of Cuba, showing Soviet nuclear missiles, their transports and tents for fueling and maintenance.][1]\r\n\r\nEdit: https://en.wikipedia.org/wiki/Cuban_Missile_Crisis\r\n\r\n\r\n [1]: https://upload.wikimedia.org/wikipedia/commons/0/0e/U2_Image_of_Cuban_Missile_Crisis.jpg", "117854": "Anyone have ideas how to detect which image was first and which one was last? Five Days is too small period to have some noticible changes in trees, buildings etc.", "117857": "Well, there may be other ways about this, but my initial intent is to coregister and reproject them. Because there's a slight oblique angle here and there, even if you figure out how to translate them into the same image space, it's likely that not all the pixels match up. \r\n\r\nYesterday I tried, for two sets, to draw all the images into the same space, but this didn't work out. I didn't have a lot of time, but it appears that the images that are closer to earth are not necessarily taken at lower altitudes, but may have different zoom levels. This can have an impact, because the camera parameters may change significantly between them.\r\n\r\nI only tried two sets, but it's already a word of warning that, irrespective of the immense detail, getting them to co-register properly may already be a daunting task.", "117968": "My current public LB score (0.29) is based on the pure ML approach (no hand labelling, no external data, no random guessing, no meta data...). I extracted features from each train set and applied to test set by Random Forest.\r\n\r\nThe strange thing is that my local CV score is not so bad, but it miserably failed on the test set. I might be doing something wrong, but it feels like there fundamental difference between them. (Is this the reason why the admin said that winning solution must be half man half machine approach?)", "118029": "[quote=ZFTurbo;117854]\r\n\r\nAnyone have ideas how to detect which image was first and which one was last? Five Days is too small period to have some noticible changes in trees, buildings etc.\r\n\r\n[/quote]\r\n\r\nUse satellite images to align all photos to true north, and to get their latitude. Use the shadow angle to calculate the time of day. Use the time of day, latitude, and shadow length to calculate solar elevation angle at noon. Sort by decreasing or increasing elevation, depending on whether the photos were taken in spring or in fall.", "118106": "Any idea how to find images on Google Earth except manual labeling?", "119018": "> Use satellite images to align all photos to true north, and to get their latitude. Use the shadow angle to calculate the time of day. Use the time of day, latitude, and shadow length to calculate solar elevation angle at noon. Sort by decreasing or increasing elevation, depending on whether the photos were taken in spring or in fall.\r\n\r\nIs that kind of approach allowed in this competition? (Assuming it was feasible)", "119254": "[quote=MattZhang;119018]\r\n\r\n> Use satellite images to align all photos to true north, and to get their latitude. Use the shadow angle to calculate the time of day. Use the time of day, latitude, and shadow length to calculate solar elevation angle at noon. Sort by decreasing or increasing elevation, depending on whether the photos were taken in spring or in fall.\r\n\r\nIs that kind of approach allowed in this competition? (Assuming it was feasible)\r\n\r\n[/quote]\r\n\r\nI think this is a man+machine approach and is should be allowed. Do we need the true north and the exact location though? All the images are from the same region and taken at the same days. And we have a training set with true orderings. I don't know but maybe there is enough information in this setup.", "119260": "| Do we need the true north and the exact location though?\r\n\r\nMy suggestion was only half-serious. To check the feasibility of the approach, I just used an online solar elevation calculator to calculate the elevation at noon today and tomorrow in San Diego. I got 74.25 and 74.51 degrees. If I did my math correctly, this means the length of a shadow of a 10 meter long vertical object changes by 0.05 meters per day.", "119435": "I think it's a interesting one. It'll need a novel idea to make any real progress I think.", "119898": "[quote=khyh;117968]\r\n\r\nThe strange thing is that my local CV score is not so bad, but it miserably failed on the test set. I might be doing something wrong, but it feels like there fundamental difference between them. (Is this the reason why the admin said that winning solution must be half man half machine approach?)\r\n\r\n[/quote]\r\n\r\nI had a local 0.55 CV yet it got crushed to 0.07 in public LB (I did not even look at the picture sets - and I had tied ranks I did not fix too, which complicated the issue of interpreting it due to Spearman rank correlation as it is clearly not explained how ties are handled). Looking at the images directly it seems there are a lot of different types of pictures to predict on. We do not have many themes given in the train set, which is quite a pain (building? water? where are forests? nature? some types of \"strange\" building, erm?). It might be possible it could predict a bit on the themes given in the train set, but failed really hard on the unknown thematics provided in the test set.\r\n\r\nP.S: I took 5 random sets in the train set and predicted manually at 4 of them perfectly. Hand labeling correction in this competition seem to be potentially much better than ML alone, have to test that on more sets (but it's a pain to order them 1 by 1).\r\n\r\n* Providing a ML solution: Draper will be interested in, \"pre-cooked\".\r\n* Providing a hand-classified solution: Draper will be really much interested at the framework you created to order the pictures, and they may find a way to translate the problem into a more accurate ML problem to solve by their own means (the framework being usually the preliminary/precursor of a larger research).\r\n\r\nAlso, it's possible to reduce the problem into a pairwise ranking problem: for each pair in the set, order the pictures by pair. Once you have all the pair rules, order the whole 5 pictures of the set by rule precedence (i.e the stronger rules supersedes the weaker rules in case of rule conflicts).", "119919": "Laurae, \r\nMay I ask on what criteria you managed to predict the 4 sets out of 5 ? what are you looking for ? I thought the change in the orientation of the shade and the length of the shade will be the key given the train set but that give me headaches for most of the pictures in the test set.\r\n\r\nMaybe a manual classification into residential, highways, river, industry, harbor, forest could help when aggregated with the dataset such as the one built by the1owl script to feed a ML classifier.\r\n\r\nI finally understood the \"registration\" show by Ben Kamphaus in his amazing script but I haven't figure how to resize the all set into the overlapping area yet. Do you think we can avoid re-sizing of the overlapping area for each set of picture in this challenge ?", "119929": "[quote=eagle4;119919]\r\n\r\n@ Laurae, \r\nMay I ask on what criteria you managed to predict the 4 sets out of 5 ? what are you looking for ? I thought the change in the orientation of the shade and the length of the shade will be the key given the train set but that give me headaches for most of the pictures in the test set.\r\n\r\n[/quote]\r\n\r\nFirst, I used registration alignment (to align using an affine transformation) then elastic registration (to scale using a proximity transformation) to make sure everything is well aligned. Then, as each images are \"closely\" aligned on every object (you have many ways to do that, it requires to use a keypoint detector such as KAZE, Harris corner detection, etc.), I focus on the very details of the each image by swapping images very quickly so my eyes can catch up potential strange differences (remember \"strange\": for instance if cars disappear in the middle of a street, it's perfectly normal). Once I spotted an area with major detail changes, I look for the same area for the five pictures.\r\n\r\nAssuming we have only two pictures (A and B) with that detail change, we can conclude either one or more of the following:\r\n\r\n* A > B\r\n* B < A\r\n* A > B or B < A\r\n* A = 1 or B = 1\r\n* A = 5 or B = 5\r\n* A = 1 or B = 1 or A = 5 or B = 5\r\n\r\nAnd I do this until I have enough rules to conclude about the complete order of the pictures.\r\n\r\nI tried to use the shade length/orientation, I found it can swap two or three pictures (due to unperfect alignment/scaling).\r\n\r\nI had issues like the following when reproducing the theoretical Z-axis (scale = order of the picture), example:\r\n\r\n![enter image description here][1]\r\n\r\nUsing only the shade, I would have predicted 4, 1, 5, 2, 3, which is wrong (and gives -0.1 Spearman rank correlation).\r\n\r\nThe major issue is finding the right details to find the right order of pictures, because sometimes it's either:\r\n\r\n* A, B, C, D, E\r\n* E, D, C, B, A\r\n\r\nAlso, I am assuming the pictures are not taken at the exact time of day each (is it the case?). And one cloudy/windy weather would be enough to dampen the usage of shades (there are pictures with severe clouds or wind).\r\n\r\nBy the way, how do you deal with 3099x2329 pictures? Do you crop 1 horizontal pixel from 3100x2329 pictures?\r\n\r\n [1]: http://i.imgur.com/80Oc8DP.png", "119998": "Re ABCDE vs EDCBA: this is especially annoying given the scoring algorithm, since EDCBA results in -1.0 even though logically you have the correct sequence." }, "source": "meta" }