It means that dark space under car became noisy mess.\nThat's why 0.999 is unreachable.\nDo you have better images?
", "messages": [ { "id": 215870, "postDate": "2017-08-23T13:22:41.050Z", "content": "It means that dark space under car became noisy mess.\nThat's why 0.999 is unreachable.\nDo you have better images?
", "rawMarkdown": "It means that dark space under car became noisy mess.\nThat's why 0.999 is unreachable.\nDo you have better images?", "votes": 18 }, { "id": 215950, "postDate": "2017-08-23T18:52:43.083Z", "content": "I was able to see this issue too:
\n\n$ identify -verbose test/0004d4463b50_01.jpg | grep Quality\n Quality: 55\nThey should have been at least 70-75, but it looks like 55 was introduced in one of my processing steps. I'll discuss this issue with Kaggle.
\n\nThanks for bringing this to our attention and sorry for the artifacts!
", "rawMarkdown": "I was able to see this issue too:\n\n $ identify -verbose test/0004d4463b50_01.jpg | grep Quality\n Quality: 55\n\nThey should have been at least 70-75, but it looks like 55 was introduced in one of my processing steps. I'll discuss this issue with Kaggle.\n\nThanks for bringing this to our attention and sorry for the artifacts!", "votes": 3, "replies": [ { "id": 215970, "postDate": "2017-08-23T20:07:46.030Z", "content": "Please, consider providing us with larger images (7500x5000) as well. You've mentioned before that this is possible. We can handle the rest of processing on our own. At least this way we'll be the only ones to blame.
", "rawMarkdown": "Please, consider providing us with larger images (7500x5000) as well. You've mentioned before that this is possible. We can handle the rest of processing on our own. At least this way we'll be the only ones to blame." }, { "id": 215974, "postDate": "2017-08-23T20:20:24.217Z", "content": "I believe that you cannot change data now within runnig competition.
\n\nBut you may create another one with better data :-D
", "rawMarkdown": "I believe that you cannot change data now within runnig competition.\n\nBut you may create another one with better data :-D", "votes": 4 }, { "id": 215975, "postDate": "2017-08-23T20:30:15.057Z", "content": "Hi Nick\nI do not think it is an option to give a new data as we are already to deep in this\nBut any way it is a good point to publish the unresized dataset after we finish this one and moreover find out any mistakes to prevent them in a Future for Carvana as model ll be implemented in a production pipeline
", "rawMarkdown": "Hi Nick\nI do not think it is an option to give a new data as we are already to deep in this\nBut any way it is a good point to publish the unresized dataset after we finish this one and moreover find out any mistakes to prevent them in a Future for Carvana as model ll be implemented in a production pipeline" }, { "id": 215984, "postDate": "2017-08-23T20:51:40.727Z", "content": "I think new data is perfectly fine at this stage of competition. It is just halfway through and there are a plenty of time to retrain any models on new data and try something new too.
\n\nI'm totally in if there will be new unprocessed 7500*5000 files 😃
", "rawMarkdown": "I think new data is perfectly fine at this stage of competition. It is just halfway through and there are a plenty of time to retrain any models on new data and try something new too. \n\nI'm totally in if there will be new unprocessed 7500*5000 files 😃", "votes": 2 }, { "id": 215986, "postDate": "2017-08-23T21:03:44.113Z", "content": "Adding higher quality images alongside the originals at the same resolution shouldn't cause anyone any problems.
\n\nChanging the resolution is a different story. As much as I think it could be helpful, at this stage of the competition, it may not be fair to make such a substantial change.
", "rawMarkdown": "Adding higher quality images alongside the originals at the same resolution shouldn't cause anyone any problems.\n\nChanging the resolution is a different story. As much as I think it could be helpful, at this stage of the competition, it may not be fair to make such a substantial change.", "votes": 5 }, { "id": 215987, "postDate": "2017-08-23T21:15:39.737Z", "content": "Brian,
\n\nCan you please clarify - have you used same images for masking?\nI see that many mask errors could be explained by those jpeg artifacts.\nThus why I'm against new HQ images.
", "rawMarkdown": "Brian,\n\nCan you please clarify - have you used same images for masking?\nI see that many mask errors could be explained by those jpeg artifacts.\nThus why I'm against new HQ images." }, { "id": 215992, "postDate": "2017-08-23T21:28:49.843Z", "content": "If the mask labels were produced from the same quality images provided to us I doubt having higher quality images available will provide a significant benefit.
", "rawMarkdown": "If the mask labels were produced from the same quality images provided to us I doubt having higher quality images available will provide a significant benefit.", "votes": 2 }, { "id": 216006, "postDate": "2017-08-23T23:31:45.977Z", "content": "@Integral Can you give an example of a mask error that appears to be caused by JPG artifacts? The mask errors I have seen have generally been inconsistent decisions of whether to follow a path, masking areas that shouldn't be (between spokes in the wheels), and some careless misses of areas that would be visible if close attention was paid (e.g. https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37229#213076).
\n\nI'm not sure about the quality settings used for the masks. If they were not the same files, they were at least the same resolution.
", "rawMarkdown": "@Integral Can you give an example of a mask error that appears to be caused by JPG artifacts? The mask errors I have seen have generally been inconsistent decisions of whether to follow a path, masking areas that shouldn't be (between spokes in the wheels), and some careless misses of areas that would be visible if close attention was paid (e.g. [https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37229#213076][1]).\n\nI'm not sure about the quality settings used for the masks. If they were not the same files, they were at least the same resolution.\n\n [1]: https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37229#213076" }, { "id": 216007, "postDate": "2017-08-23T23:39:41.053Z", "content": "@Peter It's a valid point, especially if it comes down to whoever can best predict human errors that were caused by JPG artifacts. However, there is a case to be made that the vast majority of human masks are accurate despite JPG artifacts and higher quality images can lead to more accurate predictions.
", "rawMarkdown": "@Peter It's a valid point, especially if it comes down to whoever can best predict human errors that were caused by JPG artifacts. However, there is a case to be made that the vast majority of human masks are accurate *despite* JPG artifacts and higher quality images can lead to more accurate predictions.", "votes": 1 }, { "id": 216118, "postDate": "2017-08-24T11:35:45.077Z", "content": "@Brain,\nI haven't said \"appears\", I said \"seems\".
\n\nIt's difficult to bring you exact example, but I noticed that many errors are within wheel arches. I'm not speaking about only spokes, but all stuff within arches. This part of image is quite dark and thus full of jpeg artifacts. Even without those artifacts its difficult to understand where is a car. With artifacts it sometimes became impossible
", "rawMarkdown": "@Brain,\nI haven't said \"appears\", I said \"seems\".\n\nIt's difficult to bring you exact example, but I noticed that many errors are within wheel arches. I'm not speaking about only spokes, but all stuff within arches. This part of image is quite dark and thus full of jpeg artifacts. Even without those artifacts its difficult to understand where is a car. With artifacts it sometimes became impossible " }, { "id": 216145, "postDate": "2017-08-24T13:49:51.423Z", "content": "This would surely be the best approach. A v2 dataset at the same resolution (1918x1280) with higher quality compression from the original images.\nSurely, releasing the 7500x5000 images when the competition ends would be even better!
\n\nBTW: waiting for an official confirmation about the leaderboard scores not updating if the rounded score remains the same)
", "rawMarkdown": "This would surely be the best approach. A v2 dataset at the same resolution (1918x1280) with higher quality compression from the original images.\nSurely, releasing the 7500x5000 images when the competition ends would be even better!\n\nBTW: waiting for an official confirmation about the leaderboard scores not updating if the rounded score remains the same)" }, { "id": 216149, "postDate": "2017-08-24T14:19:57.247Z", "content": "in the thread below it has been confirmed that there is a problem
", "rawMarkdown": "in the thread below it has been confirmed that there is a problem", "votes": -2 }, { "id": 216589, "postDate": "2017-08-26T19:13:56.140Z", "content": "@Brian so in general, the spokes of the wheels should be included in the mask?
", "rawMarkdown": "@Brian so in general, the spokes of the wheels should be included in the mask?" }, { "id": 217465, "postDate": "2017-08-30T17:01:21.780Z", "content": "\n\n\n@Brian so in general, the spokes of the wheels should be included in\n the mask?
\n
We don't want them to be included, but the scores will be based on the test masks which likely include them with the same frequency as the training set. For the competition, the answer is in the training data.
", "rawMarkdown": "> @Brian so in general, the spokes of the wheels should be included in\n> the mask?\n\nWe don't *want* them to be included, but the scores will be based on the test masks which likely include them with the same frequency as the training set. For the competition, the answer is in the training data." } ] }, { "id": 215886, "postDate": "2017-08-23T14:56:13.580Z", "content": "These are the images we have, I guess we will have to deal with these :)\nPreprocessing is an option..
", "rawMarkdown": "These are the images we have, I guess we will have to deal with these :)\nPreprocessing is an option..", "votes": 1, "replies": [ { "id": 215905, "postDate": "2017-08-23T16:14:50.350Z", "content": "I understand your point.\nI just want admins to understand that with better images there could be better results.\nI mean they could spend their money wisely
", "rawMarkdown": "I understand your point.\nI just want admins to understand that with better images there could be better results.\nI mean they could spend their money wisely" }, { "id": 215935, "postDate": "2017-08-23T18:06:53.220Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 215950, "author_name": "Brian Shaler", "author_url": "", "post_date": "2017-08-23T18:52:43.083000", "content": "I was able to see this issue too:
\n\n$ identify -verbose test/0004d4463b50_01.jpg | grep Quality\n Quality: 55\nThey should have been at least 70-75, but it looks like 55 was introduced in one of my processing steps. I'll discuss this issue with Kaggle.
\n\nThanks for bringing this to our attention and sorry for the artifacts!
", "votes": 3, "replies": [ { "id": 215970, "author_name": "Nikolay Shebanov", "author_url": "", "post_date": "2017-08-23T20:07:46.030000", "content": "Please, consider providing us with larger images (7500x5000) as well. You've mentioned before that this is possible. We can handle the rest of processing on our own. At least this way we'll be the only ones to blame.
", "votes": 0, "replies": [] }, { "id": 215974, "author_name": "Integral", "author_url": "", "post_date": "2017-08-23T20:20:24.217000", "content": "I believe that you cannot change data now within runnig competition.
\n\nBut you may create another one with better data :-D
", "votes": 4, "replies": [] }, { "id": 215975, "author_name": "ValeriyBabushkin", "author_url": "", "post_date": "2017-08-23T20:30:15.057000", "content": "Hi Nick\nI do not think it is an option to give a new data as we are already to deep in this\nBut any way it is a good point to publish the unresized dataset after we finish this one and moreover find out any mistakes to prevent them in a Future for Carvana as model ll be implemented in a production pipeline
", "votes": 0, "replies": [] }, { "id": 215984, "author_name": "Cut Onion", "author_url": "", "post_date": "2017-08-23T20:51:40.727000", "content": "I think new data is perfectly fine at this stage of competition. It is just halfway through and there are a plenty of time to retrain any models on new data and try something new too.
\n\nI'm totally in if there will be new unprocessed 7500*5000 files 😃
", "votes": 2, "replies": [] }, { "id": 215986, "author_name": "Brian Shaler", "author_url": "", "post_date": "2017-08-23T21:03:44.113000", "content": "Adding higher quality images alongside the originals at the same resolution shouldn't cause anyone any problems.
\n\nChanging the resolution is a different story. As much as I think it could be helpful, at this stage of the competition, it may not be fair to make such a substantial change.
", "votes": 5, "replies": [] }, { "id": 215987, "author_name": "Integral", "author_url": "", "post_date": "2017-08-23T21:15:39.737000", "content": "Brian,
\n\nCan you please clarify - have you used same images for masking?\nI see that many mask errors could be explained by those jpeg artifacts.\nThus why I'm against new HQ images.
", "votes": 0, "replies": [] }, { "id": 215992, "author_name": "Petros Giannakopoulos", "author_url": "", "post_date": "2017-08-23T21:28:49.843000", "content": "If the mask labels were produced from the same quality images provided to us I doubt having higher quality images available will provide a significant benefit.
", "votes": 2, "replies": [] }, { "id": 216006, "author_name": "Brian Shaler", "author_url": "", "post_date": "2017-08-23T23:31:45.977000", "content": "@Integral Can you give an example of a mask error that appears to be caused by JPG artifacts? The mask errors I have seen have generally been inconsistent decisions of whether to follow a path, masking areas that shouldn't be (between spokes in the wheels), and some careless misses of areas that would be visible if close attention was paid (e.g. https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37229#213076).
\n\nI'm not sure about the quality settings used for the masks. If they were not the same files, they were at least the same resolution.
", "votes": 0, "replies": [] }, { "id": 216007, "author_name": "Brian Shaler", "author_url": "", "post_date": "2017-08-23T23:39:41.053000", "content": "@Peter It's a valid point, especially if it comes down to whoever can best predict human errors that were caused by JPG artifacts. However, there is a case to be made that the vast majority of human masks are accurate despite JPG artifacts and higher quality images can lead to more accurate predictions.
", "votes": 1, "replies": [] }, { "id": 216118, "author_name": "Integral", "author_url": "", "post_date": "2017-08-24T11:35:45.077000", "content": "@Brain,\nI haven't said \"appears\", I said \"seems\".
\n\nIt's difficult to bring you exact example, but I noticed that many errors are within wheel arches. I'm not speaking about only spokes, but all stuff within arches. This part of image is quite dark and thus full of jpeg artifacts. Even without those artifacts its difficult to understand where is a car. With artifacts it sometimes became impossible
", "votes": 0, "replies": [] }, { "id": 216145, "author_name": "Davide Boschetto", "author_url": "", "post_date": "2017-08-24T13:49:51.423000", "content": "This would surely be the best approach. A v2 dataset at the same resolution (1918x1280) with higher quality compression from the original images.\nSurely, releasing the 7500x5000 images when the competition ends would be even better!
\n\nBTW: waiting for an official confirmation about the leaderboard scores not updating if the rounded score remains the same)
", "votes": 0, "replies": [] }, { "id": 216149, "author_name": "ValeriyBabushkin", "author_url": "", "post_date": "2017-08-24T14:19:57.247000", "content": "in the thread below it has been confirmed that there is a problem
", "votes": -2, "replies": [] }, { "id": 216589, "author_name": "Adam Hart", "author_url": "", "post_date": "2017-08-26T19:13:56.140000", "content": "@Brian so in general, the spokes of the wheels should be included in the mask?
", "votes": 0, "replies": [] }, { "id": 217465, "author_name": "Brian Shaler", "author_url": "", "post_date": "2017-08-30T17:01:21.780000", "content": "\n\n\n@Brian so in general, the spokes of the wheels should be included in\n the mask?
\n
We don't want them to be included, but the scores will be based on the test masks which likely include them with the same frequency as the training set. For the competition, the answer is in the training data.
", "votes": 0, "replies": [] } ] }, { "id": 215886, "author_name": "Davide Boschetto", "author_url": "", "post_date": "2017-08-23T14:56:13.580000", "content": "These are the images we have, I guess we will have to deal with these :)\nPreprocessing is an option..
", "votes": 1, "replies": [ { "id": 215905, "author_name": "Integral", "author_url": "", "post_date": "2017-08-23T16:14:50.350000", "content": "I understand your point.\nI just want admins to understand that with better images there could be better results.\nI mean they could spend their money wisely
", "votes": 0, "replies": [] }, { "id": 215935, "author_name": "", "author_url": "", "post_date": "2017-08-23T18:06:53.220000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "215870": "It means that dark space under car became noisy mess.\nThat's why 0.999 is unreachable.\nDo you have better images?", "215950": "I was able to see this issue too:\n\n $ identify -verbose test/0004d4463b50_01.jpg | grep Quality\n Quality: 55\n\nThey should have been at least 70-75, but it looks like 55 was introduced in one of my processing steps. I'll discuss this issue with Kaggle.\n\nThanks for bringing this to our attention and sorry for the artifacts!", "215886": "These are the images we have, I guess we will have to deal with these :)\nPreprocessing is an option.." } }