![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3512350%2Fcb2d221a8db8f795b462fa7e88b838bd%2FLeakHighlighted.png?generation=1572224200794654&alt=media\")
I was checking http://apolloscape.auto/car_instance.html dataset and looks like it has the same data but in slightly different format: it has a bit more meta-information, train images are of higher size on disk, but visually quality looks the same on one pair I compared (171206_034636094_Camera_5 vs ID_0e6f0cc36), also has sample scripts for visualization: https://github.com/ApolloScapeAuto/dataset-api/tree/master/car_instance (didn't check them). \nSo the train part could potentially be useful because it has a bit more meta and sample rendering and evaluation scripts (although evaluation looks different?).
Now the tricky bit is the test part: first, test part there is only 1042 images, instead of 2022 images in kaggle test. But these test images have a bit more meta (at least in the file name, like \"171206_040306841_Camera_5.jpg\"), have bigger size (not sure about quality), and if there is overlap between kaggle test and appolo test, they could potentially be useful. I didn't spot overlap after a quick look, but didn't check properly yet. Public leaderboard has only 9%, so potentially there could be overlap with a private test as well. Also some kaggle test images seem to have really really poor quality - not sure if appolo test could be helpful.
I was checking http://apolloscape.auto/car_instance.html dataset and looks like it has the same data but in slightly different format: it has a bit more meta-information, train images are of higher size on disk, but visually quality looks the same on one pair I compared (171206_034636094_Camera_5 vs ID_0e6f0cc36), also has sample scripts for visualization: https://github.com/ApolloScapeAuto/dataset-api/tree/master/car_instance (didn't check them). \nSo the train part could potentially be useful because it has a bit more meta and sample rendering and evaluation scripts (although evaluation looks different?).
Now the tricky bit is the test part: first, test part there is only 1042 images, instead of 2022 images in kaggle test. But these test images have a bit more meta (at least in the file name, like \"171206_040306841_Camera_5.jpg\"), have bigger size (not sure about quality), and if there is overlap between kaggle test and appolo test, they could potentially be useful. I didn't spot overlap after a quick look, but didn't check properly yet. Public leaderboard has only 9%, so potentially there could be overlap with a private test as well. Also some kaggle test images seem to have really really poor quality - not sure if appolo test could be helpful.
Please don't be a leak
", "rawMarkdown": "Please don't be a leak", "votes": null }, { "id": "659605", "postDate": "10/28/2019 01:05:46", "content": "Unfortunately, this looks like it may be a leak. It appears at least one of the images in the ApolloScape dataset is in this competition's test set. The image I'm talking about is ID_8ce2b0c7d.jpg. Here is a screenshot showing what I'm talking about:\n
I haven't downloaded the ApolloScape dataset or done a thorough search yet. I found this one example because a bank in the ApolloScape readme looked awfully familiar.
", "rawMarkdown": "Unfortunately, this looks like it may be a leak. It appears at least one of the images in the ApolloScape dataset is in this competition's test set. The image I'm talking about is ID_8ce2b0c7d.jpg. Here is a screenshot showing what I'm talking about:\n\n\nI haven't downloaded the ApolloScape dataset or done a thorough search yet. I found this one example because a bank in [the ApolloScape readme](https://github.com/ApolloScapeAuto/dataset-api) looked *awfully* familiar.", "votes": null }, { "id": "659647", "postDate": "10/28/2019 02:38:46", "content": "The overlap is reasonable since 'Apollo' is the name of Baidu's autonomous vehicle technology platform.
", "rawMarkdown": "The overlap is reasonable since 'Apollo' is the name of Baidu's autonomous vehicle technology platform.", "votes": null }, { "id": "660489", "postDate": "10/29/2019 08:03:37", "content": "Another related dataset is http://apolloscape.auto/scene.html - I hope test images don't overlap with this one.
\n\nAlso it's interesting that in test set, we have images from the same rides as in train, just different clips (some just 10 meters away it seems), and given that a lot of cars are stationary (parked), that could allow some tricks.
\n\nAlso it's really strange that test images were clearly post-processed, as if some set of random color augmentations were applied to them to make them look a bit worse.
", "rawMarkdown": "Another related dataset is http://apolloscape.auto/scene.html - I hope test images don't overlap with this one.\n\nAlso it's interesting that in test set, we have images from the same rides as in train, just different clips (some just 10 meters away it seems), and given that a lot of cars are stationary (parked), that could allow some tricks.\n\nAlso it's really strange that test images were clearly post-processed, as if some set of random color augmentations were applied to them to make them look a bit worse.", "votes": null }, { "id": "676730", "postDate": "11/19/2019 13:18:44", "content": "Seems that kaggle test is augmented apollo test.![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2860610%2Ffa02b0f337bc1bc34eb22f78634fc375%2FID_caa0d1bbc.jpg?generation=1574169665663385&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2860610%2F36060dd88bbaa8eb2818e09c28426f4d%2F180116_062537506_Camera_5.jpg?generation=1574169691144517&alt=media\")
Also I found that kaggle test set overlaps with sample set (3d_car_instance_sample.tar.gz) from Apollo\n(left - test, right - apollo)\n \nID_4051e4f23.jpg 180116_054006274_Camera_5.jpg\nID_f89d8e6cd.jpg 180116_053958924_Camera_5.jpg\nID_429b79cb8.jpg 180116_054112533_Camera_5.jpg\nID_9086e1d64.jpg 180116_054132978_Camera_5.jpg\nID_56052aca4.jpg 180116_054053484_Camera_5.jpg\nID_0b8753e02.jpg 180116_054047134_Camera_5.jpg\n
ID_caa0d1bbc.jpg 180116_062537506_Camera_5.jpg \ncorrespondingly
\n", "rawMarkdown": "`ID_caa0d1bbc.jpg 180116_062537506_Camera_5.jpg` \ncorrespondingly\n> Seems that kaggle test is augmented apollo test", "votes": null }, { "id": "684648", "postDate": "11/30/2019 07:04:30", "content": "Seems that kaggle test is augmented apollo test
\n
Thank you for sharing this information.\nLooking at these images carefully, ID_caa0d1bbc.jpg is not only flipped but also cropped!\nCropped image should have different camera parameters which we do not know.\nI'm wondering how can we predict world coordinates x,y,z, not knowing camera parameters.
", "rawMarkdown": "Thank you for sharing this information.\nLooking at these images carefully, ID_caa0d1bbc.jpg is not only flipped but also cropped!\nCropped image should have different camera parameters which we do not know.\nI'm wondering how can we predict world coordinates x,y,z, not knowing camera parameters.", "votes": null }, { "id": "698154", "postDate": "12/18/2019 21:37:22", "content": "It seems ApolloScape rotation annotation has different coordinate frame. This competition, data description says rotation values is relative to camera, but how about Apollo Scape? I still cannot render it...
", "rawMarkdown": "It seems ApolloScape rotation annotation has different coordinate frame. This competition, data description says rotation values is relative to camera, but how about Apollo Scape? I still cannot render it...", "votes": null }, { "id": "698518", "postDate": "12/19/2019 10:42:12", "content": "I encountered the same problem, this transformation helped somehow\nif abs(pose[1]) > math.pi / 2:\n pose[0] = -pose[0]\nbut car models are not perfectly aligned
Thanks for comment, but I believe this transformation should be done by changing coordinate frame.\nI'm still working on this...
", "rawMarkdown": "Thanks for comment, but I believe this transformation should be done by changing coordinate frame.\nI'm still working on this...", "votes": null }, { "id": "720059", "postDate": "01/16/2020 05:35:58", "content": "@bamps53 \nHave you figured out how to render it? I'm also working on it but get stuck
\n\nExample from Kaggle csv:\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1025042%2F84d8429d66af6b25d46eab9fa875eece%2F1.png?generation=1579151985824450&alt=media\")
Example from ApolloScape\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1025042%2F2f1899610ef675e4fcb32393b34ddaa9%2F2.png?generation=1579151985728049&alt=media\")
@ccchang801023 At first you need to swap yaw and pitch because somehow host mixed up. And some of the annotations in ApolloScape dataset seem collapsed. You need to omit them.
", "rawMarkdown": "ccchang801023 At first you need to swap yaw and pitch because somehow host mixed up. And some of the annotations in ApolloScape dataset seem collapsed. You need to omit them.", "votes": null }, { "id": "721774", "postDate": "01/17/2020 17:21:04", "content": "I use the apollo data to train, but the result is bad... Did you use the dataset to train?👀
", "rawMarkdown": "I use the apollo data to train, but the result is bad... Did you use the dataset to train?👀", "votes": null }, { "id": "721777", "postDate": "01/17/2020 17:24:20", "content": "Thank for replying, yes I also swapped the yaw and pitch, and I'll check the collapsed data later. but what about the (translation&rotation)transformation? Are they calculated under same coordination frame?
", "rawMarkdown": "Thank for replying, yes I also swapped the yaw and pitch, and I'll check the collapsed data later. but what about the (translation&rotation)transformation? Are they calculated under same coordination frame?", "votes": null } ], "comments": [ { "id": 659539, "author_name": "ilu000", "author_url": "", "post_date": "10/27/2019 21:32:32", "content": "Please don't be a leak
", "votes": null, "replies": [] }, { "id": 659605, "author_name": "jsday96", "author_url": "", "post_date": "10/28/2019 01:05:46", "content": "Unfortunately, this looks like it may be a leak. It appears at least one of the images in the ApolloScape dataset is in this competition's test set. The image I'm talking about is ID_8ce2b0c7d.jpg. Here is a screenshot showing what I'm talking about:\n
I haven't downloaded the ApolloScape dataset or done a thorough search yet. I found this one example because a bank in the ApolloScape readme looked awfully familiar.
", "votes": null, "replies": [] }, { "id": 659647, "author_name": "naivelamb", "author_url": "", "post_date": "10/28/2019 02:38:46", "content": "The overlap is reasonable since 'Apollo' is the name of Baidu's autonomous vehicle technology platform.
", "votes": null, "replies": [] }, { "id": 660489, "author_name": "lopuhin", "author_url": "", "post_date": "10/29/2019 08:03:37", "content": "Another related dataset is http://apolloscape.auto/scene.html - I hope test images don't overlap with this one.
\n\nAlso it's interesting that in test set, we have images from the same rides as in train, just different clips (some just 10 meters away it seems), and given that a lot of cars are stationary (parked), that could allow some tricks.
\n\nAlso it's really strange that test images were clearly post-processed, as if some set of random color augmentations were applied to them to make them look a bit worse.
", "votes": null, "replies": [] }, { "id": 676730, "author_name": "mikeedi", "author_url": "", "post_date": "11/19/2019 13:18:44", "content": "Seems that kaggle test is augmented apollo test.\n
Also I found that kaggle test set overlaps with sample set (3d_car_instance_sample.tar.gz) from Apollo\n(left - test, right - apollo)\n \nID_4051e4f23.jpg 180116_054006274_Camera_5.jpg\nID_f89d8e6cd.jpg 180116_053958924_Camera_5.jpg\nID_429b79cb8.jpg 180116_054112533_Camera_5.jpg\nID_9086e1d64.jpg 180116_054132978_Camera_5.jpg\nID_56052aca4.jpg 180116_054053484_Camera_5.jpg\nID_0b8753e02.jpg 180116_054047134_Camera_5.jpg\n
ID_caa0d1bbc.jpg 180116_062537506_Camera_5.jpg \ncorrespondingly
\n", "votes": null, "replies": [] }, { "id": 684648, "author_name": "its7171", "author_url": "", "post_date": "11/30/2019 07:04:30", "content": "Seems that kaggle test is augmented apollo test
\n
Thank you for sharing this information.\nLooking at these images carefully, ID_caa0d1bbc.jpg is not only flipped but also cropped!\nCropped image should have different camera parameters which we do not know.\nI'm wondering how can we predict world coordinates x,y,z, not knowing camera parameters.
", "votes": null, "replies": [] } ] }, { "id": 698154, "author_name": "bamps53", "author_url": "", "post_date": "12/18/2019 21:37:22", "content": "It seems ApolloScape rotation annotation has different coordinate frame. This competition, data description says rotation values is relative to camera, but how about Apollo Scape? I still cannot render it...
", "votes": null, "replies": [ { "id": 698518, "author_name": "kruntuid", "author_url": "", "post_date": "12/19/2019 10:42:12", "content": "I encountered the same problem, this transformation helped somehow\nif abs(pose[1]) > math.pi / 2:\n pose[0] = -pose[0]\nbut car models are not perfectly aligned
Thanks for comment, but I believe this transformation should be done by changing coordinate frame.\nI'm still working on this...
", "votes": null, "replies": [] }, { "id": 720059, "author_name": "ccchang801023", "author_url": "", "post_date": "01/16/2020 05:35:58", "content": "@bamps53 \nHave you figured out how to render it? I'm also working on it but get stuck
\n\nExample from Kaggle csv:\n
Example from ApolloScape\n
@ccchang801023 At first you need to swap yaw and pitch because somehow host mixed up. And some of the annotations in ApolloScape dataset seem collapsed. You need to omit them.
", "votes": null, "replies": [] }, { "id": 721774, "author_name": "guobaozi", "author_url": "", "post_date": "01/17/2020 17:21:04", "content": "I use the apollo data to train, but the result is bad... Did you use the dataset to train?👀
", "votes": null, "replies": [] }, { "id": 721777, "author_name": "ccchang801023", "author_url": "", "post_date": "01/17/2020 17:24:20", "content": "Thank for replying, yes I also swapped the yaw and pitch, and I'll check the collapsed data later. but what about the (translation&rotation)transformation? Are they calculated under same coordination frame?
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "659273": "I was checking http://apolloscape.auto/car_instance.html dataset and looks like it has the same data but in slightly different format: it has a bit more meta-information, train images are of higher size on disk, but visually quality looks the same on one pair I compared (`171206_034636094_Camera_5` vs `ID_0e6f0cc36`), also has sample scripts for visualization: https://github.com/ApolloScapeAuto/dataset-api/tree/master/car_instance (didn't check them). \nSo the train part could potentially be useful because it has a bit more meta and sample rendering and evaluation scripts (although evaluation looks different?).\n\nNow the tricky bit is the test part: first, test part there is only 1042 images, instead of 2022 images in kaggle test. But these test images have a bit more meta (at least in the file name, like \"`171206_040306841_Camera_5.jpg`\"), have bigger size (not sure about quality), and if there is overlap between kaggle test and appolo test, they could potentially be useful. I didn't spot overlap after a quick look, but didn't check properly yet. Public leaderboard has only 9%, so potentially there could be overlap with a private test as well. Also some kaggle test images seem to have really really poor quality - not sure if appolo test could be helpful.", "659539": "Please don't be a leak", "659605": "Unfortunately, this looks like it may be a leak. It appears at least one of the images in the ApolloScape dataset is in this competition's test set. The image I'm talking about is ID_8ce2b0c7d.jpg. Here is a screenshot showing what I'm talking about:\n\n\nI haven't downloaded the ApolloScape dataset or done a thorough search yet. I found this one example because a bank in [the ApolloScape readme](https://github.com/ApolloScapeAuto/dataset-api) looked *awfully* familiar.", "659647": "The overlap is reasonable since 'Apollo' is the name of Baidu's autonomous vehicle technology platform.", "660489": "Another related dataset is http://apolloscape.auto/scene.html - I hope test images don't overlap with this one.\n\nAlso it's interesting that in test set, we have images from the same rides as in train, just different clips (some just 10 meters away it seems), and given that a lot of cars are stationary (parked), that could allow some tricks.\n\nAlso it's really strange that test images were clearly post-processed, as if some set of random color augmentations were applied to them to make them look a bit worse.", "676730": "Seems that kaggle test is augmented apollo test.\n", "676740": "Also I found that kaggle test set overlaps with sample set (3d_car_instance_sample.tar.gz) from Apollo\n(left - test, right - apollo)\n ```\nID_4051e4f23.jpg 180116_054006274_Camera_5.jpg\nID_f89d8e6cd.jpg 180116_053958924_Camera_5.jpg\nID_429b79cb8.jpg 180116_054112533_Camera_5.jpg\nID_9086e1d64.jpg 180116_054132978_Camera_5.jpg\nID_56052aca4.jpg 180116_054053484_Camera_5.jpg\nID_0b8753e02.jpg 180116_054047134_Camera_5.jpg\n```", "676747": "`ID_caa0d1bbc.jpg 180116_062537506_Camera_5.jpg` \ncorrespondingly\n> Seems that kaggle test is augmented apollo test", "684648": "Thank you for sharing this information.\nLooking at these images carefully, ID_caa0d1bbc.jpg is not only flipped but also cropped!\nCropped image should have different camera parameters which we do not know.\nI'm wondering how can we predict world coordinates x,y,z, not knowing camera parameters.", "698154": "It seems ApolloScape rotation annotation has different coordinate frame. This competition, data description says rotation values is relative to camera, but how about Apollo Scape? I still cannot render it...", "698518": "I encountered the same problem, this transformation helped somehow\n` if abs(pose[1]) > math.pi / 2:\n pose[0] = -pose[0]`\nbut car models are not perfectly aligned", "698628": "Thanks for comment, but I believe this transformation should be done by changing coordinate frame.\nI'm still working on this...", "720059": "bamps53 \nHave you figured out how to render it? I'm also working on it but get stuck\n\n Example from Kaggle csv:\n\n\nExample from ApolloScape\n", "721162": "ccchang801023 At first you need to swap yaw and pitch because somehow host mixed up. And some of the annotations in ApolloScape dataset seem collapsed. You need to omit them.", "721774": "I use the apollo data to train, but the result is bad... Did you use the dataset to train?👀", "721777": "Thank for replying, yes I also swapped the yaw and pitch, and I'll check the collapsed data later. but what about the (translation&rotation)transformation? Are they calculated under same coordination frame?" }, "source": "meta" }