After browsing through the training dataset, I noticed that you can do a pretty good job in categorizing the images just by looking at how the pictures were taken instead of focusing on the chronological changes that took place in the images. For example, days 1 and 2 are always taken at a low altitude which makes them appear zoomed in compared to days 3, 4, and 5. Also, the images for days 4 and 5 are always taken at the same angle (as in the slope of straight lines in the image) which makes it easy to differentiate them from day 3. View all images of a few sets next to each other and you'll see what I am talking about.
\n\nUpdate:
\n\nIt appears that pictures in the test dataset are taken at different times than the pictures in the training dataset so the pattern I described probably won't help to classify the images in the training set. I am not sure if there is still some pattern in how the pictures were taken in the test dataset so it might have been a false alarm.
", "messages": [ { "id": 117689, "postDate": "2016-04-30T08:05:22.057Z", "content": "After browsing through the training dataset, I noticed that you can do a pretty good job in categorizing the images just by looking at how the pictures were taken instead of focusing on the chronological changes that took place in the images. For example, days 1 and 2 are always taken at a low altitude which makes them appear zoomed in compared to days 3, 4, and 5. Also, the images for days 4 and 5 are always taken at the same angle (as in the slope of straight lines in the image) which makes it easy to differentiate them from day 3. View all images of a few sets next to each other and you'll see what I am talking about.
\n\nUpdate:
\n\nIt appears that pictures in the test dataset are taken at different times than the pictures in the training dataset so the pattern I described probably won't help to classify the images in the training set. I am not sure if there is still some pattern in how the pictures were taken in the test dataset so it might have been a false alarm.
", "rawMarkdown": "After browsing through the training dataset, I noticed that you can do a pretty good job in categorizing the images just by looking at how the pictures were taken instead of focusing on the chronological changes that took place in the images. For example, days 1 and 2 are always taken at a low altitude which makes them appear zoomed in compared to days 3, 4, and 5. Also, the images for days 4 and 5 are always taken at the same angle (as in the slope of straight lines in the image) which makes it easy to differentiate them from day 3. View all images of a few sets next to each other and you'll see what I am talking about.\r\n\r\n**Update**:\r\n\r\nIt appears that pictures in the test dataset are taken at different times than the pictures in the training dataset so the pattern I described probably won't help to classify the images in the training set. I am not sure if there is still some pattern in how the pictures were taken in the test dataset so it might have been a false alarm.\r\n", "votes": 9 }, { "id": 117709, "postDate": "2016-04-30T13:34:54.247Z", "content": "There are many problems with this dataset. There are a few sets that are overlapping, solving one means solving all of them. Even if they are not overlapping you can tell they were taken in the same area, looking at the white balance, altitude, time of day, shadows you can find a correspondence between them. Again solving one set means solving all of them. Moreover if you have 10-20 sets that are linked together then you can assume that at least one set will be in the public LB fold, so you can easily check the correct sequence by checking the public LB score. There are only 274 sets and 60 days of competition, manually solving 4-5 sets a day is more than possible, it's inviting.
\n\nThis competition has little, if anything, to do with machine learning. It feels like it's a recruiting challenge for NSA or NRO. Just check Draper's wiki page: https://en.wikipedia.org/wiki/Charles_Stark_Draper_Laboratory
\n\nEDIT: Draper and NRO http://www.draper.com/news/jana-schwartz-awarded-national-reconnaissance-office-fellowship
", "rawMarkdown": "There are many problems with this dataset. There are a few sets that are overlapping, solving one means solving all of them. Even if they are not overlapping you can tell they were taken in the same area, looking at the white balance, altitude, time of day, shadows you can find a correspondence between them. Again solving one set means solving all of them. Moreover if you have 10-20 sets that are linked together then you can assume that at least one set will be in the public LB fold, so you can easily check the correct sequence by checking the public LB score. There are only 274 sets and 60 days of competition, manually solving 4-5 sets a day is more than possible, it's inviting.\r\n\r\nThis competition has little, if anything, to do with machine learning. It feels like it's a recruiting challenge for NSA or NRO. Just check Draper's wiki page: https://en.wikipedia.org/wiki/Charles_Stark_Draper_Laboratory\r\n\r\nEDIT: Draper and NRO http://www.draper.com/news/jana-schwartz-awarded-national-reconnaissance-office-fellowship", "votes": 8 }, { "id": 119829, "postDate": "2016-05-12T23:55:57.113Z", "content": "95% on 1 submission does seem suspicious
", "rawMarkdown": "95% on 1 submission does seem suspicious", "votes": 5 }, { "id": 119805, "postDate": "2016-05-12T20:47:56.290Z", "content": "We have been in contact with one of the leading groups and do not currently have reason to believe there is leakage. For now, the competition continues as-is (n.b. as with all Kaggle competitions, timelines and datasets can change if leakage rears its ugly head).
", "rawMarkdown": "We have been in contact with one of the leading groups and do not currently have reason to believe there is leakage. For now, the competition continues as-is (n.b. as with all Kaggle competitions, timelines and datasets can change if leakage rears its ugly head).", "votes": 3 }, { "id": 117707, "postDate": "2016-04-30T13:24:47.693Z", "content": "I must admit that I had similar reservations after browsing through the images. Given this data has been gathered by aircraft there's a risk of leakage from inferring flight path and altitude. There are many spatially overlapping image sets which I would have thought also create leakage risk.
\n\nLet's hope that the competition admins have been clever and anticipated this.
\n\nIf there is leakage this could be a very short competition!
", "rawMarkdown": "I must admit that I had similar reservations after browsing through the images. Given this data has been gathered by aircraft there's a risk of leakage from inferring flight path and altitude. There are many spatially overlapping image sets which I would have thought also create leakage risk.\r\n\r\nLet's hope that the competition admins have been clever and anticipated this. \r\n\r\nIf there is leakage this could be a very short competition!", "votes": 3 }, { "id": 117972, "postDate": "2016-05-02T05:08:24.407Z", "content": "Teams are limited to size 5 in this competition
", "rawMarkdown": "Teams are limited to size 5 in this competition", "votes": 4 }, { "id": 117956, "postDate": "2016-05-02T02:42:51.467Z", "content": "[quote=Radu Stoicescu;117709]
\n\nThere are only 274 sets and 60 days of competition, manually solving 4-5 sets a day is more than possible, it's inviting.
\n\n[/quote]
\n\nOr we could all join a single team, massively parallelize the job, and be done with the whole thing in an hour. ;) I'd be happy with my share of $500. :D
", "rawMarkdown": "[quote=Radu Stoicescu;117709]\r\n\r\nThere are only 274 sets and 60 days of competition, manually solving 4-5 sets a day is more than possible, it's inviting.\r\n\r\n[/quote]\r\n\r\nOr we could all join a single team, massively parallelize the job, and be done with the whole thing in an hour. ;) I'd be happy with my share of $500. :D", "votes": 4 }, { "id": 120031, "postDate": "2016-05-14T18:53:33.630Z", "content": "[quote=Chippy;119921]
\n\nI thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...
\n\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3
\n\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?
\n\n[/quote]
\n\nGoogle Earth seems to have more up-to-date imagery and time-stamping than Google maps. The Google Earth image taken on 3/Feb/2016 at Set59 location is very similar to the Draper image. The Googlemap image is closer to the Google Earth image dated Jan 2014.
\n\nSo Google Earth images and the ability to go back in time may be helpful as @Tyler Vigen suggested.
", "rawMarkdown": "[quote=Chippy;119921]\r\n\r\nI thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...\r\n\r\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3\r\n\r\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?\r\n\r\n[/quote]\r\n\r\n\r\nGoogle Earth seems to have more up-to-date imagery and time-stamping than Google maps. The Google Earth image taken on 3/Feb/2016 at Set59 location is very similar to the Draper image. The Googlemap image is closer to the Google Earth image dated Jan 2014.\r\n\r\nSo **Google Earth** images and the ability to go back in time may be helpful as @Tyler Vigen suggested.\r\n", "votes": 1 }, { "id": 119864, "postDate": "2016-05-13T10:22:17.127Z", "content": "Just interested: Does anyone here participate because he only wants to win money? For me it's fun and knowing my own limits. Therefore hand labeling for me isn't an option.
", "rawMarkdown": "Just interested: Does anyone here participate because he only wants to win money? For me it's fun and knowing my own limits. Therefore hand labeling for me isn't an option. ", "votes": 1 }, { "id": 119780, "postDate": "2016-05-12T18:28:09.037Z", "content": "You could. Have you looked at the images though? I certainly couldn't order most of them reliably!
", "rawMarkdown": "You could. Have you looked at the images though? I certainly couldn't order most of them reliably!", "votes": 1 }, { "id": 119863, "postDate": "2016-05-13T10:18:44.973Z", "content": "@the1owl\nI think sarcasm has no place in a $75k hand labelling competition. Stay focused and keep on labelling.
", "rawMarkdown": "@the1owl\r\nI think sarcasm has no place in a $75k hand labelling competition. Stay focused and keep on labelling.", "votes": -1 }, { "id": 119984, "postDate": "2016-05-14T10:35:57.467Z", "content": "[quote=Chippy;119970]
\n\nThanks @Tyler Vigen. Are there any Image sets and web links you can share to illustrate?
\n\n[/quote]\nThis is just a technique I used in a similar previous project. I'm still working on finding perfect pairs to generalize to this set! I'll let you know if I come up with an obvious grouping.
", "rawMarkdown": "[quote=Chippy;119970]\r\n\r\nThanks @Tyler Vigen. Are there any Image sets and web links you can share to illustrate?\r\n\r\n[/quote]\r\nThis is just a technique I used in a similar previous project. I'm still working on finding perfect pairs to generalize to this set! I'll let you know if I come up with an obvious grouping." }, { "id": 119970, "postDate": "2016-05-14T06:50:05.110Z", "content": "Thanks @Tyler Vigen. Are there any Image sets and web links you can share to illustrate?
", "rawMarkdown": "Thanks @Tyler Vigen. Are there any Image sets and web links you can share to illustrate?" }, { "id": 119952, "postDate": "2016-05-14T00:27:52.063Z", "content": "[quote=Chippy;119921]
\n\nI thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...
\n\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3
\n\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?
\n\n[/quote]\nGoogle has cached images from both the month before and the month after the training photos were taken, depending on where you're looking. Combine that with any number of other aerial photo providers and geotagged Flickr photographs with the right datetime stamp and you'd be surprised how far you can get.
\n\nFlickr ground feature photos might not be ideal for machine learning, but there are plenty of technically generalizable techniques that apply to that kind of extrapolated geospatial analysis.
", "rawMarkdown": "[quote=Chippy;119921]\r\n\r\nI thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...\r\n\r\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3\r\n\r\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?\r\n\r\n[/quote]\r\nGoogle has cached images from both the month before and the month after the training photos were taken, depending on where you're looking. Combine that with any number of other aerial photo providers and geotagged Flickr photographs with the right datetime stamp and you'd be surprised how far you can get.\r\n\r\nFlickr ground feature photos might not be ideal for machine learning, but there are plenty of technically generalizable techniques that apply to that kind of extrapolated geospatial analysis." }, { "id": 119949, "postDate": "2016-05-13T22:58:29.377Z", "content": "[quote=Chippy;119921]
\n\nI thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...
\n\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3
\n\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?
\n\n[/quote]
\n\nIt seems that even Set59_1 is more recent than that Google Maps image. What does Google Earth historical imagery time slider gives? (instead of using Google Maps)
", "rawMarkdown": "[quote=Chippy;119921]\r\n\r\nI thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...\r\n\r\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3\r\n\r\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?\r\n\r\n[/quote]\r\n\r\nIt seems that even Set59_1 is more recent than that Google Maps image. What does Google Earth historical imagery time slider gives? (instead of using Google Maps)" }, { "id": 119921, "postDate": "2016-05-13T18:58:05.547Z", "content": "I thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...
\n\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3
\n\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?
", "rawMarkdown": "I thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...\r\n\r\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3\r\n\r\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?" }, { "id": 119917, "postDate": "2016-05-13T18:29:59.600Z", "content": "[quote=Jared Turkewitz;119804]
\n\n[quote=Paulo Pinto;119800]
\n\nI can really assure you that there is a methodology to find out the exact date and time of each image generated by satellite or aircraft , within the rules of the challenge and can work with error almost zero , for any set of non- manipulated images so almost automatic ( with little manual intervention) . So I believe that the winner of this challenge will reach the index 1 .
\n\n[/quote]
\n\nIs it using external data? For example do you think the leak is basically using some version of cached google maps?
\n\n[/quote]
\n\nYes.
", "rawMarkdown": "[quote=Jared Turkewitz;119804]\r\n\r\n[quote=Paulo Pinto;119800]\r\n\r\nI can really assure you that there is a methodology to find out the exact date and time of each image generated by satellite or aircraft , within the rules of the challenge and can work with error almost zero , for any set of non- manipulated images so almost automatic ( with little manual intervention) . So I believe that the winner of this challenge will reach the index 1 .\r\n\r\n[/quote]\r\n\r\nIs it using external data? For example do you think the leak is basically using some version of cached google maps?\r\n\r\n[/quote]\r\n\r\nYes." }, { "id": 119811, "postDate": "2016-05-12T21:02:21.613Z", "content": "[quote=William Cukierski;119805]
\n\nWe have been in contact with one of the leading groups and do not currently have reason to believe there is leakage. For now, the competition continues as-is (n.b. as with all Kaggle competitions, timelines and datasets can change if leakage rears its ugly head).
\n\n[/quote]\nThis is good to know, but it doesn't seem like there are very many submissions that would qualify as resulting from leakage in this competition given the broadness of the rules and inputs.
", "rawMarkdown": "[quote=William Cukierski;119805]\r\n\r\nWe have been in contact with one of the leading groups and do not currently have reason to believe there is leakage. For now, the competition continues as-is (n.b. as with all Kaggle competitions, timelines and datasets can change if leakage rears its ugly head).\r\n\r\n[/quote]\r\nThis is good to know, but it doesn't seem like there are very many submissions that would qualify as resulting from leakage in this competition given the broadness of the rules and inputs." }, { "id": 119804, "postDate": "2016-05-12T20:46:45.467Z", "content": "[quote=Paulo Pinto;119800]
\n\nI can really assure you that there is a methodology to find out the exact date and time of each image generated by satellite or aircraft , within the rules of the challenge and can work with error almost zero , for any set of non- manipulated images so almost automatic ( with little manual intervention) . So I believe that the winner of this challenge will reach the index 1 .
\n\n[/quote]
\n\nIs it using external data? For example do you think the leak is basically using some version of cached google maps?
", "rawMarkdown": "[quote=Paulo Pinto;119800]\r\n\r\nI can really assure you that there is a methodology to find out the exact date and time of each image generated by satellite or aircraft , within the rules of the challenge and can work with error almost zero , for any set of non- manipulated images so almost automatic ( with little manual intervention) . So I believe that the winner of this challenge will reach the index 1 .\r\n\r\n[/quote]\r\n\r\nIs it using external data? For example do you think the leak is basically using some version of cached google maps?" }, { "id": 119800, "postDate": "2016-05-12T20:32:03.337Z", "content": "I can really assure you that there is a methodology to find out the exact date and time of each image generated by satellite or aircraft , within the rules of the challenge and can work with error almost zero , for any set of non- manipulated images so almost automatic ( with little manual intervention) . So I believe that the winner of this challenge will reach the index 1 .
", "rawMarkdown": "I can really assure you that there is a methodology to find out the exact date and time of each image generated by satellite or aircraft , within the rules of the challenge and can work with error almost zero , for any set of non- manipulated images so almost automatic ( with little manual intervention) . So I believe that the winner of this challenge will reach the index 1 ." }, { "id": 119777, "postDate": "2016-05-12T18:22:18.397Z", "content": "OK, that makes sense. Thanks for the reply.
\n\nFor hand labeling wouldn't you just look at each group of five, and choose order you think, enter it into spreadsheet & submit?
", "rawMarkdown": "OK, that makes sense. Thanks for the reply.\r\n\r\n For hand labeling wouldn't you just look at each group of five, and choose order you think, enter it into spreadsheet & submit?" }, { "id": 119774, "postDate": "2016-05-12T18:05:02.567Z", "content": "A leak would be a way of classifying the photos order without temporal features in the photo. Noticing that one days photos were always the same scale relative the the others for example. Basically some method that works great with this data set but which would fail on any other.
\n\nHand labelling them all would be fine. Still got to work out how to do that though!
", "rawMarkdown": "A leak would be a way of classifying the photos order without temporal features in the photo. Noticing that one days photos were always the same scale relative the the others for example. Basically some method that works great with this data set but which would fail on any other.\r\n\r\nHand labelling them all would be fine. Still got to work out how to do that though!\r\n" }, { "id": 119772, "postDate": "2016-05-12T17:47:27.143Z", "content": "Hi, I'm kinda new to Kaggle.\n What is meant by a "leak" in this context? Seems it's not necessarily a bad thing.?.
\n\nBut, it is possible that it was all done by hand; from other forum posts saying that hand labeling is allowed on both training & test; seems rules do not expressly forbid it. With 1370 images, 274 groups of 5, to sort, a person can do that in a few days; & its been almost two weeks since data was posted. Wonder if Kaggle/Draper will do a final check with a separate test set; it's only important if Draper wants an algorithm or is looking to hire a person, for whatever this data challenge means to them.
", "rawMarkdown": " Hi, I'm kinda new to Kaggle.\r\n What is meant by a \"leak\" in this context? Seems it's not necessarily a bad thing.?.\r\n\r\n But, it is possible that it was all done by hand; from other forum posts saying that hand labeling is allowed on both training & test; seems rules do not expressly forbid it. With 1370 images, 274 groups of 5, to sort, a person can do that in a few days; & its been almost two weeks since data was posted. Wonder if Kaggle/Draper will do a final check with a separate test set; it's only important if Draper wants an algorithm or is looking to hire a person, for whatever this data challenge means to them." }, { "id": 119755, "postDate": "2016-05-12T16:28:50.423Z", "content": "This situation is not so good for Kaggler's motivation. )
", "rawMarkdown": "This situation is not so good for Kaggler's motivation. )" }, { "id": 119752, "postDate": "2016-05-12T16:03:10.647Z", "content": "Great work by Vladimir Tomecek
\n\nIf it is a leak (and we don't know for sure) Kaggle usually ask the leader how they did it prior to the competition deadline
", "rawMarkdown": "Great work by Vladimir Tomecek \r\n\r\n***If*** it is a leak (and we don't know for sure) Kaggle usually ask the leader how they did it prior to the competition deadline\r\n" }, { "id": 119750, "postDate": "2016-05-12T15:54:23.353Z", "content": "I would bet Kaggle and Draper have no idea what the leak may be. If they reset they may never find out, so my bet is that until guys on the top of the leaderboard keep leak secret there would be no reset of the competition. If someone publishes it the reset is very probable.
", "rawMarkdown": "I would bet Kaggle and Draper have no idea what the leak may be. If they reset they may never find out, so my bet is that until guys on the top of the leaderboard keep leak secret there would be no reset of the competition. If someone publishes it the reset is very probable." }, { "id": 119747, "postDate": "2016-05-12T15:51:01.393Z", "content": "I think the best move from Kaggle would be to upload new images a week before the deadline. So everyone is able to run the code and upload the new submission. But wouldn't be able to leak it again.
", "rawMarkdown": "I think the best move from Kaggle would be to upload new images a week before the deadline. So everyone is able to run the code and upload the new submission. But wouldn't be able to leak it again. " }, { "id": 119745, "postDate": "2016-05-12T15:43:16.557Z", "content": "Any bets on a reset? Now the best is .95
", "rawMarkdown": "Any bets on a reset? Now the best is .95" }, { "id": 119733, "postDate": "2016-05-12T14:27:58.553Z", "content": "[quote=the1owl;119732]
\n\nIn image processing completions what could be considered a leak? The randomness?
\n\n[/quote]
\n\nI think a leak would be if every second image was rotated or zoomed in the same way (like in the train set :( ) Cause only a few people found it at the moment the leak is much harder :D But that kind of stuff is considered as a leak I suppose.
", "rawMarkdown": "[quote=the1owl;119732]\r\n\r\nIn image processing completions what could be considered a leak? The randomness?\r\n\r\n[/quote]\r\n\r\nI think a leak would be if every second image was rotated or zoomed in the same way (like in the train set :( ) Cause only a few people found it at the moment the leak is much harder :D But that kind of stuff is considered as a leak I suppose.\r\n" }, { "id": 119731, "postDate": "2016-05-12T14:17:46.240Z", "content": "Does anyone spend some days to do it only by hand? I've watched some images for too long now and I'm pretty sure that I can't even do this by hand. It was a nice challenge guys ;)
", "rawMarkdown": "Does anyone spend some days to do it only by hand? I've watched some images for too long now and I'm pretty sure that I can't even do this by hand. It was a nice challenge guys ;)" }, { "id": 119730, "postDate": "2016-05-12T14:15:47.273Z", "content": "[quote=Ole Kröger;119728]
\n\nHaha well we have a score of nearly 1 now :D
\n\n[/quote]
\n\n... and with only one single submission. Obviously leak found in CV already.
", "rawMarkdown": "[quote=Ole Kröger;119728]\r\n\r\nHaha well we have a score of nearly 1 now :D\r\n\r\n[/quote]\r\n\r\n... and with only one single submission. Obviously leak found in CV already." }, { "id": 119728, "postDate": "2016-05-12T14:12:29.733Z", "content": "Haha well we have a score of nearly 1 now :D
", "rawMarkdown": "Haha well we have a score of nearly 1 now :D" }, { "id": 119701, "postDate": "2016-05-12T11:06:02.260Z", "content": "[quote=Chippy;117707]
\n\n....
\n\nIf there is leakage this could be a very short competition!
\n\n[/quote]
\n\nWell done Bad Robot, so we are all wondering is this leak exploitation, hard work or a bit of both?
\n\nI've not been able to work on this much over the last week and automated leak exploitation hasn't been as easy for me as I'd first thought. So I'm still hoping that Bad Robot has done some clever data science otherwise the leaderboard will be full of perfect scores of 1 within a week...!
", "rawMarkdown": "[quote=Chippy;117707]\r\n\r\n....\r\n\r\nIf there is leakage this could be a very short competition!\r\n\r\n[/quote]\r\n\r\nWell done Bad Robot, so we are all wondering is this leak exploitation, hard work or a bit of both?\r\n\r\nI've not been able to work on this much over the last week and automated leak exploitation hasn't been as easy for me as I'd first thought. So I'm still hoping that Bad Robot has done some clever data science otherwise the leaderboard will be full of perfect scores of 1 within a week...!\r\n " }, { "id": 119695, "postDate": "2016-05-12T09:37:01.713Z", "content": "Congratulations to Bad Robot, that was fast.
", "rawMarkdown": "Congratulations to Bad Robot, that was fast." }, { "id": 119689, "postDate": "2016-05-12T09:09:50.057Z", "content": "My guess is that a sudden score of ~0.75 answers the 'is there a leak?' question.
", "rawMarkdown": "My guess is that a sudden score of ~0.75 answers the 'is there a leak?' question. " }, { "id": 119849, "postDate": "2016-05-13T05:18:46.210Z", "rawMarkdown": "", "votes": 4, "isDeleted": true }, { "id": 119732, "postDate": "2016-05-12T14:21:04.250Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 117709, "author_name": "Radu Stoicescu", "author_url": "", "post_date": "2016-04-30T13:34:54.247000", "content": "There are many problems with this dataset. There are a few sets that are overlapping, solving one means solving all of them. Even if they are not overlapping you can tell they were taken in the same area, looking at the white balance, altitude, time of day, shadows you can find a correspondence between them. Again solving one set means solving all of them. Moreover if you have 10-20 sets that are linked together then you can assume that at least one set will be in the public LB fold, so you can easily check the correct sequence by checking the public LB score. There are only 274 sets and 60 days of competition, manually solving 4-5 sets a day is more than possible, it's inviting.
\n\nThis competition has little, if anything, to do with machine learning. It feels like it's a recruiting challenge for NSA or NRO. Just check Draper's wiki page: https://en.wikipedia.org/wiki/Charles_Stark_Draper_Laboratory
\n\nEDIT: Draper and NRO http://www.draper.com/news/jana-schwartz-awarded-national-reconnaissance-office-fellowship
", "votes": 8, "replies": [] }, { "id": 119829, "author_name": "Andre L Torres", "author_url": "", "post_date": "2016-05-12T23:55:57.113000", "content": "95% on 1 submission does seem suspicious
", "votes": 5, "replies": [] }, { "id": 119805, "author_name": "Will Cukierski", "author_url": "", "post_date": "2016-05-12T20:47:56.290000", "content": "We have been in contact with one of the leading groups and do not currently have reason to believe there is leakage. For now, the competition continues as-is (n.b. as with all Kaggle competitions, timelines and datasets can change if leakage rears its ugly head).
", "votes": 3, "replies": [] }, { "id": 117707, "author_name": "Nigel Carpenter", "author_url": "", "post_date": "2016-04-30T13:24:47.693000", "content": "I must admit that I had similar reservations after browsing through the images. Given this data has been gathered by aircraft there's a risk of leakage from inferring flight path and altitude. There are many spatially overlapping image sets which I would have thought also create leakage risk.
\n\nLet's hope that the competition admins have been clever and anticipated this.
\n\nIf there is leakage this could be a very short competition!
", "votes": 3, "replies": [] }, { "id": 117972, "author_name": "Jared Turkewitz", "author_url": "", "post_date": "2016-05-02T05:08:24.407000", "content": "Teams are limited to size 5 in this competition
", "votes": 4, "replies": [] }, { "id": 117956, "author_name": "Bojan Tunguz", "author_url": "", "post_date": "2016-05-02T02:42:51.467000", "content": "[quote=Radu Stoicescu;117709]
\n\nThere are only 274 sets and 60 days of competition, manually solving 4-5 sets a day is more than possible, it's inviting.
\n\n[/quote]
\n\nOr we could all join a single team, massively parallelize the job, and be done with the whole thing in an hour. ;) I'd be happy with my share of $500. :D
", "votes": 4, "replies": [] }, { "id": 120031, "author_name": "Nigel Carpenter", "author_url": "", "post_date": "2016-05-14T18:53:33.630000", "content": "[quote=Chippy;119921]
\n\nI thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...
\n\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3
\n\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?
\n\n[/quote]
\n\nGoogle Earth seems to have more up-to-date imagery and time-stamping than Google maps. The Google Earth image taken on 3/Feb/2016 at Set59 location is very similar to the Draper image. The Googlemap image is closer to the Google Earth image dated Jan 2014.
\n\nSo Google Earth images and the ability to go back in time may be helpful as @Tyler Vigen suggested.
", "votes": 1, "replies": [] }, { "id": 119864, "author_name": "Ole Kröger", "author_url": "", "post_date": "2016-05-13T10:22:17.127000", "content": "Just interested: Does anyone here participate because he only wants to win money? For me it's fun and knowing my own limits. Therefore hand labeling for me isn't an option.
", "votes": 1, "replies": [] }, { "id": 119780, "author_name": "hairykrishna", "author_url": "", "post_date": "2016-05-12T18:28:09.037000", "content": "You could. Have you looked at the images though? I certainly couldn't order most of them reliably!
", "votes": 1, "replies": [] }, { "id": 119863, "author_name": "EnvyMyFieldsMedal", "author_url": "", "post_date": "2016-05-13T10:18:44.973000", "content": "@the1owl\nI think sarcasm has no place in a $75k hand labelling competition. Stay focused and keep on labelling.
", "votes": -1, "replies": [] }, { "id": 119984, "author_name": "Tyler Vigen", "author_url": "", "post_date": "2016-05-14T10:35:57.467000", "content": "[quote=Chippy;119970]
\n\nThanks @Tyler Vigen. Are there any Image sets and web links you can share to illustrate?
\n\n[/quote]\nThis is just a technique I used in a similar previous project. I'm still working on finding perfect pairs to generalize to this set! I'll let you know if I come up with an obvious grouping.
", "votes": 0, "replies": [] }, { "id": 119970, "author_name": "Nigel Carpenter", "author_url": "", "post_date": "2016-05-14T06:50:05.110000", "content": "Thanks @Tyler Vigen. Are there any Image sets and web links you can share to illustrate?
", "votes": 0, "replies": [] }, { "id": 119952, "author_name": "Tyler Vigen", "author_url": "", "post_date": "2016-05-14T00:27:52.063000", "content": "[quote=Chippy;119921]
\n\nI thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...
\n\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3
\n\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?
\n\n[/quote]\nGoogle has cached images from both the month before and the month after the training photos were taken, depending on where you're looking. Combine that with any number of other aerial photo providers and geotagged Flickr photographs with the right datetime stamp and you'd be surprised how far you can get.
\n\nFlickr ground feature photos might not be ideal for machine learning, but there are plenty of technically generalizable techniques that apply to that kind of extrapolated geospatial analysis.
", "votes": 0, "replies": [] }, { "id": 119949, "author_name": "Laurae", "author_url": "", "post_date": "2016-05-13T22:58:29.377000", "content": "[quote=Chippy;119921]
\n\nI thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...
\n\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3
\n\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?
\n\n[/quote]
\n\nIt seems that even Set59_1 is more recent than that Google Maps image. What does Google Earth historical imagery time slider gives? (instead of using Google Maps)
", "votes": 0, "replies": [] }, { "id": 119921, "author_name": "Nigel Carpenter", "author_url": "", "post_date": "2016-05-13T18:58:05.547000", "content": "I thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...
\n\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3
\n\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?
", "votes": 0, "replies": [] }, { "id": 119917, "author_name": "Paulo Pinto", "author_url": "", "post_date": "2016-05-13T18:29:59.600000", "content": "[quote=Jared Turkewitz;119804]
\n\n[quote=Paulo Pinto;119800]
\n\nI can really assure you that there is a methodology to find out the exact date and time of each image generated by satellite or aircraft , within the rules of the challenge and can work with error almost zero , for any set of non- manipulated images so almost automatic ( with little manual intervention) . So I believe that the winner of this challenge will reach the index 1 .
\n\n[/quote]
\n\nIs it using external data? For example do you think the leak is basically using some version of cached google maps?
\n\n[/quote]
\n\nYes.
", "votes": 0, "replies": [] }, { "id": 119811, "author_name": "Tyler Vigen", "author_url": "", "post_date": "2016-05-12T21:02:21.613000", "content": "[quote=William Cukierski;119805]
\n\nWe have been in contact with one of the leading groups and do not currently have reason to believe there is leakage. For now, the competition continues as-is (n.b. as with all Kaggle competitions, timelines and datasets can change if leakage rears its ugly head).
\n\n[/quote]\nThis is good to know, but it doesn't seem like there are very many submissions that would qualify as resulting from leakage in this competition given the broadness of the rules and inputs.
", "votes": 0, "replies": [] }, { "id": 119804, "author_name": "Jared Turkewitz", "author_url": "", "post_date": "2016-05-12T20:46:45.467000", "content": "[quote=Paulo Pinto;119800]
\n\nI can really assure you that there is a methodology to find out the exact date and time of each image generated by satellite or aircraft , within the rules of the challenge and can work with error almost zero , for any set of non- manipulated images so almost automatic ( with little manual intervention) . So I believe that the winner of this challenge will reach the index 1 .
\n\n[/quote]
\n\nIs it using external data? For example do you think the leak is basically using some version of cached google maps?
", "votes": 0, "replies": [] }, { "id": 119800, "author_name": "Paulo Pinto", "author_url": "", "post_date": "2016-05-12T20:32:03.337000", "content": "I can really assure you that there is a methodology to find out the exact date and time of each image generated by satellite or aircraft , within the rules of the challenge and can work with error almost zero , for any set of non- manipulated images so almost automatic ( with little manual intervention) . So I believe that the winner of this challenge will reach the index 1 .
", "votes": 0, "replies": [] }, { "id": 119777, "author_name": "DiscipleOfScience", "author_url": "", "post_date": "2016-05-12T18:22:18.397000", "content": "OK, that makes sense. Thanks for the reply.
\n\nFor hand labeling wouldn't you just look at each group of five, and choose order you think, enter it into spreadsheet & submit?
", "votes": 0, "replies": [] }, { "id": 119774, "author_name": "hairykrishna", "author_url": "", "post_date": "2016-05-12T18:05:02.567000", "content": "A leak would be a way of classifying the photos order without temporal features in the photo. Noticing that one days photos were always the same scale relative the the others for example. Basically some method that works great with this data set but which would fail on any other.
\n\nHand labelling them all would be fine. Still got to work out how to do that though!
", "votes": 0, "replies": [] }, { "id": 119772, "author_name": "DiscipleOfScience", "author_url": "", "post_date": "2016-05-12T17:47:27.143000", "content": "Hi, I'm kinda new to Kaggle.\n What is meant by a "leak" in this context? Seems it's not necessarily a bad thing.?.
\n\nBut, it is possible that it was all done by hand; from other forum posts saying that hand labeling is allowed on both training & test; seems rules do not expressly forbid it. With 1370 images, 274 groups of 5, to sort, a person can do that in a few days; & its been almost two weeks since data was posted. Wonder if Kaggle/Draper will do a final check with a separate test set; it's only important if Draper wants an algorithm or is looking to hire a person, for whatever this data challenge means to them.
", "votes": 0, "replies": [] }, { "id": 119755, "author_name": "ZFTurbo", "author_url": "", "post_date": "2016-05-12T16:28:50.423000", "content": "This situation is not so good for Kaggler's motivation. )
", "votes": 0, "replies": [] }, { "id": 119752, "author_name": "Scirpus", "author_url": "", "post_date": "2016-05-12T16:03:10.647000", "content": "Great work by Vladimir Tomecek
\n\nIf it is a leak (and we don't know for sure) Kaggle usually ask the leader how they did it prior to the competition deadline
", "votes": 0, "replies": [] }, { "id": 119750, "author_name": "itod", "author_url": "", "post_date": "2016-05-12T15:54:23.353000", "content": "I would bet Kaggle and Draper have no idea what the leak may be. If they reset they may never find out, so my bet is that until guys on the top of the leaderboard keep leak secret there would be no reset of the competition. If someone publishes it the reset is very probable.
", "votes": 0, "replies": [] }, { "id": 119747, "author_name": "Ole Kröger", "author_url": "", "post_date": "2016-05-12T15:51:01.393000", "content": "I think the best move from Kaggle would be to upload new images a week before the deadline. So everyone is able to run the code and upload the new submission. But wouldn't be able to leak it again.
", "votes": 0, "replies": [] }, { "id": 119745, "author_name": "Scirpus", "author_url": "", "post_date": "2016-05-12T15:43:16.557000", "content": "Any bets on a reset? Now the best is .95
", "votes": 0, "replies": [] }, { "id": 119733, "author_name": "Ole Kröger", "author_url": "", "post_date": "2016-05-12T14:27:58.553000", "content": "[quote=the1owl;119732]
\n\nIn image processing completions what could be considered a leak? The randomness?
\n\n[/quote]
\n\nI think a leak would be if every second image was rotated or zoomed in the same way (like in the train set :( ) Cause only a few people found it at the moment the leak is much harder :D But that kind of stuff is considered as a leak I suppose.
", "votes": 0, "replies": [] }, { "id": 119731, "author_name": "Ole Kröger", "author_url": "", "post_date": "2016-05-12T14:17:46.240000", "content": "Does anyone spend some days to do it only by hand? I've watched some images for too long now and I'm pretty sure that I can't even do this by hand. It was a nice challenge guys ;)
", "votes": 0, "replies": [] }, { "id": 119730, "author_name": "", "author_url": "", "post_date": "2016-05-12T14:15:47.273000", "content": "", "votes": 0, "replies": [] }, { "id": 119728, "author_name": "", "author_url": "", "post_date": "2016-05-12T14:12:29.733000", "content": "", "votes": 0, "replies": [] }, { "id": 119701, "author_name": "", "author_url": "", "post_date": "2016-05-12T11:06:02.260000", "content": "", "votes": 0, "replies": [] }, { "id": 119695, "author_name": "", "author_url": "", "post_date": "2016-05-12T09:37:01.713000", "content": "", "votes": 0, "replies": [] }, { "id": 119689, "author_name": "", "author_url": "", "post_date": "2016-05-12T09:09:50.057000", "content": "", "votes": 0, "replies": [] }, { "id": 119849, "author_name": "", "author_url": "", "post_date": "2016-05-13T05:18:46.210000", "content": "", "votes": 4, "replies": [] }, { "id": 119732, "author_name": "", "author_url": "", "post_date": "2016-05-12T14:21:04.250000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "117689": "After browsing through the training dataset, I noticed that you can do a pretty good job in categorizing the images just by looking at how the pictures were taken instead of focusing on the chronological changes that took place in the images. For example, days 1 and 2 are always taken at a low altitude which makes them appear zoomed in compared to days 3, 4, and 5. Also, the images for days 4 and 5 are always taken at the same angle (as in the slope of straight lines in the image) which makes it easy to differentiate them from day 3. View all images of a few sets next to each other and you'll see what I am talking about.\r\n\r\n**Update**:\r\n\r\nIt appears that pictures in the test dataset are taken at different times than the pictures in the training dataset so the pattern I described probably won't help to classify the images in the training set. I am not sure if there is still some pattern in how the pictures were taken in the test dataset so it might have been a false alarm.\r\n", "117709": "There are many problems with this dataset. There are a few sets that are overlapping, solving one means solving all of them. Even if they are not overlapping you can tell they were taken in the same area, looking at the white balance, altitude, time of day, shadows you can find a correspondence between them. Again solving one set means solving all of them. Moreover if you have 10-20 sets that are linked together then you can assume that at least one set will be in the public LB fold, so you can easily check the correct sequence by checking the public LB score. There are only 274 sets and 60 days of competition, manually solving 4-5 sets a day is more than possible, it's inviting.\r\n\r\nThis competition has little, if anything, to do with machine learning. It feels like it's a recruiting challenge for NSA or NRO. Just check Draper's wiki page: https://en.wikipedia.org/wiki/Charles_Stark_Draper_Laboratory\r\n\r\nEDIT: Draper and NRO http://www.draper.com/news/jana-schwartz-awarded-national-reconnaissance-office-fellowship", "119829": "95% on 1 submission does seem suspicious", "119805": "We have been in contact with one of the leading groups and do not currently have reason to believe there is leakage. For now, the competition continues as-is (n.b. as with all Kaggle competitions, timelines and datasets can change if leakage rears its ugly head).", "117707": "I must admit that I had similar reservations after browsing through the images. Given this data has been gathered by aircraft there's a risk of leakage from inferring flight path and altitude. There are many spatially overlapping image sets which I would have thought also create leakage risk.\r\n\r\nLet's hope that the competition admins have been clever and anticipated this. \r\n\r\nIf there is leakage this could be a very short competition!", "117972": "Teams are limited to size 5 in this competition", "117956": "[quote=Radu Stoicescu;117709]\r\n\r\nThere are only 274 sets and 60 days of competition, manually solving 4-5 sets a day is more than possible, it's inviting.\r\n\r\n[/quote]\r\n\r\nOr we could all join a single team, massively parallelize the job, and be done with the whole thing in an hour. ;) I'd be happy with my share of $500. :D", "120031": "[quote=Chippy;119921]\r\n\r\nI thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...\r\n\r\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3\r\n\r\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?\r\n\r\n[/quote]\r\n\r\n\r\nGoogle Earth seems to have more up-to-date imagery and time-stamping than Google maps. The Google Earth image taken on 3/Feb/2016 at Set59 location is very similar to the Draper image. The Googlemap image is closer to the Google Earth image dated Jan 2014.\r\n\r\nSo **Google Earth** images and the ability to go back in time may be helpful as @Tyler Vigen suggested.\r\n", "119864": "Just interested: Does anyone here participate because he only wants to win money? For me it's fun and knowing my own limits. Therefore hand labeling for me isn't an option. ", "119780": "You could. Have you looked at the images though? I certainly couldn't order most of them reliably!", "119863": "@the1owl\r\nI think sarcasm has no place in a $75k hand labelling competition. Stay focused and keep on labelling.", "119984": "[quote=Chippy;119970]\r\n\r\nThanks @Tyler Vigen. Are there any Image sets and web links you can share to illustrate?\r\n\r\n[/quote]\r\nThis is just a technique I used in a similar previous project. I'm still working on finding perfect pairs to generalize to this set! I'll let you know if I come up with an obvious grouping.", "119970": "Thanks @Tyler Vigen. Are there any Image sets and web links you can share to illustrate?", "119952": "[quote=Chippy;119921]\r\n\r\nI thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...\r\n\r\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3\r\n\r\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?\r\n\r\n[/quote]\r\nGoogle has cached images from both the month before and the month after the training photos were taken, depending on where you're looking. Combine that with any number of other aerial photo providers and geotagged Flickr photographs with the right datetime stamp and you'd be surprised how far you can get.\r\n\r\nFlickr ground feature photos might not be ideal for machine learning, but there are plenty of technically generalizable techniques that apply to that kind of extrapolated geospatial analysis.", "119949": "[quote=Chippy;119921]\r\n\r\nI thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...\r\n\r\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3\r\n\r\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?\r\n\r\n[/quote]\r\n\r\nIt seems that even Set59_1 is more recent than that Google Maps image. What does Google Earth historical imagery time slider gives? (instead of using Google Maps)", "119921": "I thought that these images are generally more recent that those of Google maps. Compare Set59_2 to it's googlemaps image...\r\n\r\nhttps://www.google.co.uk/maps/@33.7610227,-118.2306424,282m/data=!3m1!1e3\r\n\r\nTo me the Set59_2 image looks more recent. So I can't see how caches of Googlemaps images can help here. What do others think?", "119917": "[quote=Jared Turkewitz;119804]\r\n\r\n[quote=Paulo Pinto;119800]\r\n\r\nI can really assure you that there is a methodology to find out the exact date and time of each image generated by satellite or aircraft , within the rules of the challenge and can work with error almost zero , for any set of non- manipulated images so almost automatic ( with little manual intervention) . So I believe that the winner of this challenge will reach the index 1 .\r\n\r\n[/quote]\r\n\r\nIs it using external data? For example do you think the leak is basically using some version of cached google maps?\r\n\r\n[/quote]\r\n\r\nYes.", "119811": "[quote=William Cukierski;119805]\r\n\r\nWe have been in contact with one of the leading groups and do not currently have reason to believe there is leakage. For now, the competition continues as-is (n.b. as with all Kaggle competitions, timelines and datasets can change if leakage rears its ugly head).\r\n\r\n[/quote]\r\nThis is good to know, but it doesn't seem like there are very many submissions that would qualify as resulting from leakage in this competition given the broadness of the rules and inputs.", "119804": "[quote=Paulo Pinto;119800]\r\n\r\nI can really assure you that there is a methodology to find out the exact date and time of each image generated by satellite or aircraft , within the rules of the challenge and can work with error almost zero , for any set of non- manipulated images so almost automatic ( with little manual intervention) . So I believe that the winner of this challenge will reach the index 1 .\r\n\r\n[/quote]\r\n\r\nIs it using external data? For example do you think the leak is basically using some version of cached google maps?", "119800": "I can really assure you that there is a methodology to find out the exact date and time of each image generated by satellite or aircraft , within the rules of the challenge and can work with error almost zero , for any set of non- manipulated images so almost automatic ( with little manual intervention) . So I believe that the winner of this challenge will reach the index 1 .", "119777": "OK, that makes sense. Thanks for the reply.\r\n\r\n For hand labeling wouldn't you just look at each group of five, and choose order you think, enter it into spreadsheet & submit?", "119774": "A leak would be a way of classifying the photos order without temporal features in the photo. Noticing that one days photos were always the same scale relative the the others for example. Basically some method that works great with this data set but which would fail on any other.\r\n\r\nHand labelling them all would be fine. Still got to work out how to do that though!\r\n", "119772": " Hi, I'm kinda new to Kaggle.\r\n What is meant by a \"leak\" in this context? Seems it's not necessarily a bad thing.?.\r\n\r\n But, it is possible that it was all done by hand; from other forum posts saying that hand labeling is allowed on both training & test; seems rules do not expressly forbid it. With 1370 images, 274 groups of 5, to sort, a person can do that in a few days; & its been almost two weeks since data was posted. Wonder if Kaggle/Draper will do a final check with a separate test set; it's only important if Draper wants an algorithm or is looking to hire a person, for whatever this data challenge means to them.", "119755": "This situation is not so good for Kaggler's motivation. )", "119752": "Great work by Vladimir Tomecek \r\n\r\n***If*** it is a leak (and we don't know for sure) Kaggle usually ask the leader how they did it prior to the competition deadline\r\n", "119750": "I would bet Kaggle and Draper have no idea what the leak may be. If they reset they may never find out, so my bet is that until guys on the top of the leaderboard keep leak secret there would be no reset of the competition. If someone publishes it the reset is very probable.", "119747": "I think the best move from Kaggle would be to upload new images a week before the deadline. So everyone is able to run the code and upload the new submission. But wouldn't be able to leak it again. ", "119745": "Any bets on a reset? Now the best is .95", "119733": "[quote=the1owl;119732]\r\n\r\nIn image processing completions what could be considered a leak? The randomness?\r\n\r\n[/quote]\r\n\r\nI think a leak would be if every second image was rotated or zoomed in the same way (like in the train set :( ) Cause only a few people found it at the moment the leak is much harder :D But that kind of stuff is considered as a leak I suppose.\r\n", "119731": "Does anyone spend some days to do it only by hand? I've watched some images for too long now and I'm pretty sure that I can't even do this by hand. It was a nice challenge guys ;)", "119730": "[quote=Ole Kröger;119728]\r\n\r\nHaha well we have a score of nearly 1 now :D\r\n\r\n[/quote]\r\n\r\n... and with only one single submission. Obviously leak found in CV already.", "119728": "Haha well we have a score of nearly 1 now :D", "119701": "[quote=Chippy;117707]\r\n\r\n....\r\n\r\nIf there is leakage this could be a very short competition!\r\n\r\n[/quote]\r\n\r\nWell done Bad Robot, so we are all wondering is this leak exploitation, hard work or a bit of both?\r\n\r\nI've not been able to work on this much over the last week and automated leak exploitation hasn't been as easy for me as I'd first thought. So I'm still hoping that Bad Robot has done some clever data science otherwise the leaderboard will be full of perfect scores of 1 within a week...!\r\n ", "119695": "Congratulations to Bad Robot, that was fast.", "119689": "My guess is that a sudden score of ~0.75 answers the 'is there a leak?' question. ", "119849": "", "119732": "" } }