Competition rules say:
\n\n\nUse of external data is permitted, provided the data is freely available. If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.\n
However, I don't see an official forum thread - so I will start here. :)
\n\nI would like to use SegNet: https://github.com/alexgkendall/caffe-segnet
", "messages": [ { "id": 207526, "postDate": "2017-07-26T20:26:10.867Z", "content": "Competition rules say:
\n\n\nUse of external data is permitted, provided the data is freely available. If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.\n
However, I don't see an official forum thread - so I will start here. :)
\n\nI would like to use SegNet: https://github.com/alexgkendall/caffe-segnet
", "rawMarkdown": " Competition rules say: \n\n```\nUse of external data is permitted, provided the data is freely available. If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.\n```\n\nHowever, I don't see an official forum thread - so I will start here. :)\n\nI would like to use SegNet: https://github.com/alexgkendall/caffe-segnet\n\n", "votes": 12 }, { "id": 222334, "postDate": "2017-09-18T13:48:23.753Z", "content": "A little bit late, but we have managed to hack GTA-V and generate synthetic data!\nhttps://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781
", "rawMarkdown": "A little bit late, but we have managed to hack GTA-V and generate synthetic data!\nhttps://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781", "votes": 7, "replies": [ { "id": 222335, "postDate": "2017-09-18T13:50:11.690Z", "content": "The question to organizers:\nCould you accept this cool data? We've spent 2 weeks to generate them!
", "rawMarkdown": "The question to organizers:\nCould you accept this cool data? We've spent 2 weeks to generate them!", "votes": 1 }, { "id": 222343, "postDate": "2017-09-18T14:06:53.057Z", "content": "It's super cool! Congratulations!
\nHave you measured the improvement of using that dataset?
Sorry it's too late
", "rawMarkdown": "It's super cool! Congratulations! \nHave you measured the improvement of using that dataset?\n\nSorry it's too late\n" }, { "id": 222376, "postDate": "2017-09-18T16:08:25.160Z", "content": "That is for you to check the improvement.
\n\nAre you answering on behalf of the organizers?
", "rawMarkdown": "That is for you to check the improvement.\n\nAre you answering on behalf of the organizers?" } ] }, { "id": 222926, "postDate": "2017-09-20T15:39:53.280Z", "content": "Note to organizers: you should probably FREEZE posting new messages in this thread when the deadline hit, it would avoid any doubts
", "rawMarkdown": "Note to organizers: you should probably FREEZE posting new messages in this thread when the deadline hit, it would avoid any doubts", "votes": 3 }, { "id": 218252, "postDate": "2017-09-03T06:05:33.733Z", "content": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models
", "rawMarkdown": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models", "votes": 3 }, { "id": 215801, "postDate": "2017-08-23T06:25:58.780Z", "content": "GTA5 image and semantic label datasets: https://download.visinf.tu-darmstadt.de/data/from_games/
", "rawMarkdown": "GTA5 image and semantic label datasets: https://download.visinf.tu-darmstadt.de/data/from_games/", "votes": 3, "replies": [ { "id": 215828, "postDate": "2017-08-23T09:35:50.707Z", "content": "That is a very interesting paper.
\nMoreover it will be much easier making a mod of the game that replaces the textures of the objects by plain colors, that way they don't have to label the images.
Keras pre-trained models https://keras.io/applications/
\n\nMight also use this external dataset https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781
", "rawMarkdown": "Keras pre-trained models https://keras.io/applications/\n\nMight also use this external dataset https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781", "votes": 1 }, { "id": 208803, "postDate": "2017-07-31T05:41:40.023Z", "content": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models
", "rawMarkdown": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models", "votes": 1 }, { "id": 207528, "postDate": "2017-07-26T20:33:33.053Z", "content": "Thanks for kicking it off! I've added \"official\" to the title and pinned this thread.
", "rawMarkdown": "Thanks for kicking it off! I've added \"official\" to the title and pinned this thread.", "votes": 1, "replies": [ { "id": 222625, "postDate": "2017-09-19T13:38:12.953Z", "content": "Can you clarify the deadline for external data?
", "rawMarkdown": "Can you clarify the deadline for external data?" } ] }, { "id": 223431, "postDate": "2017-09-22T04:17:23.887Z", "content": "Using the crf-cnn pretrained model https://github.com/torrvision/crfasrnn
", "rawMarkdown": "Using the crf-cnn pretrained model https://github.com/torrvision/crfasrnn" }, { "id": 222765, "postDate": "2017-09-20T00:50:11.283Z", "content": "The usual:
\n\nI hope it's not too late! Typically the model posting dateline is by entry deadline or 1 week before the end of competition but it is worded differently here. Anyway, these models have been used in many other previous competitions and have been posted by others earlier.
", "rawMarkdown": "The usual:\n\n- PyTorch: https://github.com/pytorch/vision/tree/master/torchvision/models\n- Standard Keras models: https://keras.io/applications/\n- Various other models: https://github.com/flyyufelix/cnn_finetune\n\nI hope it's not too late! Typically the model posting dateline is by entry deadline or 1 week before the end of competition but it is worded differently here. Anyway, these models have been used in many other previous competitions and have been posted by others earlier." }, { "id": 222736, "postDate": "2017-09-19T20:12:48.717Z", "content": "Might use this external data: https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781
", "rawMarkdown": "Might use this external data: https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781", "replies": [ { "id": 223174, "postDate": "2017-09-21T10:48:55.930Z", "content": "+1
", "rawMarkdown": "+1" } ] }, { "id": 222666, "postDate": "2017-09-19T16:35:50.530Z", "content": "I just started working on this competition and have built a solution based on using the pretrained models available in the keras applications module. https://keras.io/applications/ I figured that if I could still join the competition that I would also be able to use data that other people are using; if not then what is the point of requiring that we post our data sources? However it is my reading of the rules that I will not be able to use my current solution because I did not personally post that I was using this data a week ago. However it is clearly the intent of this rule that competitors should be allowed to use the same data as each other. With this in mind it might it seems only fair that any data that has been posted to the external data board by other competitors should be allowed to be used by anyone.
\n\nBased on the wording of the rules however;
\n\n\"If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.\"
\n\nIt seems that if the data can only be used by the people that actually posted about using it prior to the 13th. Because of that I am retraining my model from scratch so as to not be using the image net weights but for future competitions it might be desirable to change the rules so that in situations like this people may still use the data that other people are using even if they didn't post about it personally before the deadline.
", "rawMarkdown": "I just started working on this competition and have built a solution based on using the pretrained models available in the keras applications module. https://keras.io/applications/ I figured that if I could still join the competition that I would also be able to use data that other people are using; if not then what is the point of requiring that we post our data sources? However it is my reading of the rules that I will not be able to use my current solution because I did not personally post that I was using this data a week ago. However it is clearly the intent of this rule that competitors should be allowed to use the same data as each other. With this in mind it might it seems only fair that any data that has been posted to the external data board by other competitors should be allowed to be used by anyone. \n\nBased on the wording of the rules however;\n\n \"If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.\" \n\nIt seems that if the data can only be used by the people that actually posted about using it prior to the 13th. Because of that I am retraining my model from scratch so as to not be using the image net weights but for future competitions it might be desirable to change the rules so that in situations like this people may still use the data that other people are using even if they didn't post about it personally before the deadline.", "replies": [ { "id": 222730, "postDate": "2017-09-19T19:41:25.827Z", "content": "Well noone enforces this rule if you are not a price winner I think. And winning a price this late in the competition seems unlikely, so I would not worry to much!
", "rawMarkdown": "Well noone enforces this rule if you are not a price winner I think. And winning a price this late in the competition seems unlikely, so I would not worry to much!", "votes": 1 }, { "id": 222739, "postDate": "2017-09-19T20:29:45.903Z", "content": "While I agree with you in principle. It just isn't as much fun to work on the problem if the idea of coming out at the top of the leader board causes me distress instead of pleasure.
", "rawMarkdown": "While I agree with you in principle. It just isn't as much fun to work on the problem if the idea of coming out at the top of the leader board causes me distress instead of pleasure. " }, { "id": 222875, "postDate": "2017-09-20T12:31:45.563Z", "content": "@ Tim, I think you are just nitpicking. For me it seems OK to use data if somebody posted it here one week before the deadline. Why would everyone duplicate the same information here? \nSo I'm sure you shouldn't worry about using Imagenet pretrained weights.
", "rawMarkdown": "@ Tim, I think you are just nitpicking. For me it seems OK to use data if somebody posted it here one week before the deadline. Why would everyone duplicate the same information here? \nSo I'm sure you shouldn't worry about using Imagenet pretrained weights.", "votes": 3 }, { "id": 222879, "postDate": "2017-09-20T12:42:25.653Z", "content": "However it is important, because if someone in a good position like you is using external data that is valuable information for all the other participants.
\n\nThat's why the rules are very clear as posted previously
\n\n\n\n\nUse of external data is permitted, provided the data is freely available. If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.
\n
@Artem.Sanakoev: I did not disagree with you at all :)
", "rawMarkdown": "@Artem.Sanakoev: I did not disagree with you at all :)", "votes": 1 }, { "id": 222979, "postDate": "2017-09-20T19:05:31.483Z", "content": "Didn't notice that there are 2 Tims in the thread:)\nI meant @Tim Anderton :)
", "rawMarkdown": "Didn't notice that there are 2 Tims in the thread:)\nI meant @Tim Anderton :)", "votes": 1 }, { "id": 223188, "postDate": "2017-09-21T11:22:56.040Z", "content": "Oh wow, me neither. :D
", "rawMarkdown": "Oh wow, me neither. :D" } ] }, { "id": 222661, "postDate": "2017-09-19T16:14:49.407Z", "content": "ResNet50
", "rawMarkdown": "ResNet50" }, { "id": 222514, "postDate": "2017-09-19T04:55:09.040Z", "content": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models
", "rawMarkdown": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models" }, { "id": 222337, "postDate": "2017-09-18T13:53:35.287Z", "content": "Pretrained models from https://keras.io/applications/
", "rawMarkdown": "Pretrained models from https://keras.io/applications/" }, { "id": 222332, "postDate": "2017-09-18T13:45:09.993Z", "content": "I might use extra data from this source: https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781
", "rawMarkdown": "I might use extra data from this source: https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781", "replies": [ { "id": 222333, "postDate": "2017-09-18T13:47:30.603Z", "content": "Looks like it's too late. It's less than 1 week till merger deadline
", "rawMarkdown": "Looks like it's too late. It's less than 1 week till merger deadline" }, { "id": 222395, "postDate": "2017-09-18T17:26:12.147Z", "content": "I believe the deadline is actually less than 1 week prior to entry deadline, not the merger deadline...so there is still time :)
", "rawMarkdown": "I believe the deadline is actually less than 1 week prior to entry deadline, not the merger deadline...so there is still time :)" }, { "id": 222401, "postDate": "2017-09-18T17:38:32.173Z", "content": "As far as I understand the last day for publishing external data was 13 September (Entry deadline is 20 September)
\n\nhttps://www.kaggle.com/c/carvana-image-masking-challenge/rules
\n\n\n", "rawMarkdown": "As far as I understand the last day for publishing external data was 13 September (Entry deadline is 20 September)\n\nhttps://www.kaggle.com/c/carvana-image-masking-challenge/rules\n> Use of external data is permitted, provided the data is freely available. If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.\n", "votes": 2 }, { "id": 222411, "postDate": "2017-09-18T18:37:41.677Z", "content": "Use of external data is permitted, provided the data is freely available. If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.
\n
OK my bad I thought \"entry deadline\" meant \"submission entry\", but gtk that entry in this context means entry into the competition....If it had just said \"merger deadline\" I think it would have been a lot more clear :)
", "rawMarkdown": "OK my bad I thought \"entry deadline\" meant \"submission entry\", but gtk that entry in this context means entry into the competition....If it had just said \"merger deadline\" I think it would have been a lot more clear :)", "votes": 1 } ] }, { "id": 222255, "postDate": "2017-09-18T07:21:57.277Z", "content": "Pre-trained models from https://github.com/facebook/fb.resnet.torch/tree/master/pretrained and \nhttps://github.com/fchollet/deep-learning-models
", "rawMarkdown": "Pre-trained models from https://github.com/facebook/fb.resnet.torch/tree/master/pretrained and \nhttps://github.com/fchollet/deep-learning-models" }, { "id": 222203, "postDate": "2017-09-18T03:11:04.867Z", "content": "Pretrained model: VGG-16 with BN, Resnet-34, Resnet-50
", "rawMarkdown": "Pretrained model: VGG-16 with BN, Resnet-34, Resnet-50" }, { "id": 221446, "postDate": "2017-09-15T09:59:28.913Z", "content": "Resnet-152 weights for Keras https://gist.github.com/flyyufelix/7e2eafb149f72f4d38dd661882c554a6
", "rawMarkdown": "Resnet-152 weights for Keras https://gist.github.com/flyyufelix/7e2eafb149f72f4d38dd661882c554a6" }, { "id": 220845, "postDate": "2017-09-13T08:53:44.147Z", "content": "Pre-trained models from tensorflow object detection api https://github.com/tensorflow/models/blob/master/object_detection/g3doc/detection_model_zoo.md
", "rawMarkdown": "Pre-trained models from tensorflow object detection api https://github.com/tensorflow/models/blob/master/object_detection/g3doc/detection_model_zoo.md" }, { "id": 220612, "postDate": "2017-09-12T17:12:57.310Z", "content": "Models posted by Marvin Teichmann: \nhttps://github.com/MarvinTeichmann/KittiSeg\nhttps://github.com/MarvinTeichmann/tensorflow-fcn
\n\nPre-trained model data for these models: \nftp://mi.eng.cam.ac.uk/pub/mttt2/models/
\n\nThese models are TensorFlow implementations and examples of the network architecture described in \"Fully Convolutional Networks for Semantic Segmentation\" by Jonathan Long, Evan Shelhamer, and Trevor Darrell from UC Berkeley: https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf
", "rawMarkdown": "Models posted by Marvin Teichmann: \nhttps://github.com/MarvinTeichmann/KittiSeg\nhttps://github.com/MarvinTeichmann/tensorflow-fcn\n\nPre-trained model data for these models: \nftp://mi.eng.cam.ac.uk/pub/mttt2/models/\n\nThese models are TensorFlow implementations and examples of the network architecture described in \"Fully Convolutional Networks for Semantic Segmentation\" by Jonathan Long, Evan Shelhamer, and Trevor Darrell from UC Berkeley: https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf" }, { "id": 220509, "postDate": "2017-09-12T11:20:19.183Z", "content": "Pre-trained models from Deeplab https://bitbucket.org/deeplab/deeplab-public/
", "rawMarkdown": "Pre-trained models from Deeplab https://bitbucket.org/deeplab/deeplab-public/" }, { "id": 219711, "postDate": "2017-09-09T08:56:05.580Z", "content": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models
", "rawMarkdown": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models" }, { "id": 219470, "postDate": "2017-09-08T09:56:11.723Z", "content": "Pre-trained mobilenet model/weights: https://github.com/fchollet/keras/blob/master/keras/applications/mobilenet.py
", "rawMarkdown": "Pre-trained mobilenet model/weights: https://github.com/fchollet/keras/blob/master/keras/applications/mobilenet.py" }, { "id": 218193, "postDate": "2017-09-02T21:39:39.887Z", "content": "In the rule it was said external data is permitted. If I google for images of cars online and added them to the training set. Is it allowed?
", "rawMarkdown": "In the rule it was said external data is permitted. If I google for images of cars online and added them to the training set. Is it allowed? ", "replies": [ { "id": 218254, "postDate": "2017-09-03T06:22:46.263Z", "content": "In other challenges it was allowed but you had to publish the links to the images and to the masks.
", "rawMarkdown": "In other challenges it was allowed but you had to publish the links to the images and to the masks." } ] }, { "id": 215807, "postDate": "2017-08-23T06:48:21.073Z", "content": "I was using the FCN for segmentation however it is relatively slow. Thanks to your idea and I decided to try the Segnet.
", "rawMarkdown": "I was using the FCN for segmentation however it is relatively slow. Thanks to your idea and I decided to try the Segnet." }, { "id": 207563, "postDate": "2017-07-26T22:19:11.427Z", "content": "Never had a chance to try SegNet and I'm waiting to see what kind of models are going to use, this competition is quite exciting.
\n\nHappy Kaggling!
", "rawMarkdown": "Never had a chance to try SegNet and I'm waiting to see what kind of models are going to use, this competition is quite exciting.\n\nHappy Kaggling!" }, { "id": 207531, "postDate": "2017-07-26T20:40:20.770Z", "content": "\"Creative Commons Attribution-NonCommercial 4.0 International License\"
\n\nCan it be used as it is intended only for non-commercial use?
", "rawMarkdown": "\"Creative Commons Attribution-NonCommercial 4.0 International License\"\n\nCan it be used as it is intended only for non-commercial use?", "replies": [ { "id": 207549, "postDate": "2017-07-26T21:14:20.860Z", "content": "By my interpretation, I would assume not: https://wiki.creativecommons.org/wiki/NonCommercial_interpretation
", "rawMarkdown": "By my interpretation, I would assume not: https://wiki.creativecommons.org/wiki/NonCommercial_interpretation" } ] }, { "id": 219823, "postDate": "2017-09-09T19:27:40.720Z", "rawMarkdown": "", "isDeleted": true }, { "id": 218816, "postDate": "2017-09-06T00:27:15.723Z", "rawMarkdown": "", "isDeleted": true }, { "id": 218130, "postDate": "2017-09-02T13:01:43.540Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 222334, "author_name": "Renat Bashirov", "author_url": "", "post_date": "2017-09-18T13:48:23.753000", "content": "A little bit late, but we have managed to hack GTA-V and generate synthetic data!\nhttps://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781
", "votes": 7, "replies": [ { "id": 222335, "author_name": "Renat Bashirov", "author_url": "", "post_date": "2017-09-18T13:50:11.690000", "content": "The question to organizers:\nCould you accept this cool data? We've spent 2 weeks to generate them!
", "votes": 1, "replies": [] }, { "id": 222343, "author_name": "Guillermo Barbadillo", "author_url": "", "post_date": "2017-09-18T14:06:53.057000", "content": "It's super cool! Congratulations!
\nHave you measured the improvement of using that dataset?
Sorry it's too late
", "votes": 0, "replies": [] }, { "id": 222376, "author_name": "Renat Bashirov", "author_url": "", "post_date": "2017-09-18T16:08:25.160000", "content": "That is for you to check the improvement.
\n\nAre you answering on behalf of the organizers?
", "votes": 0, "replies": [] } ] }, { "id": 222926, "author_name": "Andrés Miguel Torrubia Sáez", "author_url": "", "post_date": "2017-09-20T15:39:53.280000", "content": "Note to organizers: you should probably FREEZE posting new messages in this thread when the deadline hit, it would avoid any doubts
", "votes": 3, "replies": [] }, { "id": 218252, "author_name": "albu", "author_url": "", "post_date": "2017-09-03T06:05:33.733000", "content": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models
", "votes": 3, "replies": [] }, { "id": 215801, "author_name": "Eugene", "author_url": "", "post_date": "2017-08-23T06:25:58.780000", "content": "GTA5 image and semantic label datasets: https://download.visinf.tu-darmstadt.de/data/from_games/
", "votes": 3, "replies": [ { "id": 215828, "author_name": "Guillermo Barbadillo", "author_url": "", "post_date": "2017-08-23T09:35:50.707000", "content": "That is a very interesting paper.
\nMoreover it will be much easier making a mod of the game that replaces the textures of the objects by plain colors, that way they don't have to label the images.
Keras pre-trained models https://keras.io/applications/
\n\nMight also use this external dataset https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781
", "votes": 1, "replies": [] }, { "id": 208803, "author_name": "Aleksei Tiulpin", "author_url": "", "post_date": "2017-07-31T05:41:40.023000", "content": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models
", "votes": 1, "replies": [] }, { "id": 207528, "author_name": "Will Cukierski", "author_url": "", "post_date": "2017-07-26T20:33:33.053000", "content": "Thanks for kicking it off! I've added \"official\" to the title and pinned this thread.
", "votes": 1, "replies": [ { "id": 222625, "author_name": "Andrés Miguel Torrubia Sáez", "author_url": "", "post_date": "2017-09-19T13:38:12.953000", "content": "Can you clarify the deadline for external data?
", "votes": 0, "replies": [] } ] }, { "id": 223431, "author_name": "rmehta1987", "author_url": "", "post_date": "2017-09-22T04:17:23.887000", "content": "Using the crf-cnn pretrained model https://github.com/torrvision/crfasrnn
", "votes": 0, "replies": [] }, { "id": 222765, "author_name": "Kyle Lee", "author_url": "", "post_date": "2017-09-20T00:50:11.283000", "content": "The usual:
\n\nI hope it's not too late! Typically the model posting dateline is by entry deadline or 1 week before the end of competition but it is worded differently here. Anyway, these models have been used in many other previous competitions and have been posted by others earlier.
", "votes": 0, "replies": [] }, { "id": 222736, "author_name": "Harun Gunaydin", "author_url": "", "post_date": "2017-09-19T20:12:48.717000", "content": "Might use this external data: https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781
", "votes": 0, "replies": [ { "id": 223174, "author_name": "buzzlight", "author_url": "", "post_date": "2017-09-21T10:48:55.930000", "content": "+1
", "votes": 0, "replies": [] } ] }, { "id": 222666, "author_name": "Tim Anderton", "author_url": "", "post_date": "2017-09-19T16:35:50.530000", "content": "I just started working on this competition and have built a solution based on using the pretrained models available in the keras applications module. https://keras.io/applications/ I figured that if I could still join the competition that I would also be able to use data that other people are using; if not then what is the point of requiring that we post our data sources? However it is my reading of the rules that I will not be able to use my current solution because I did not personally post that I was using this data a week ago. However it is clearly the intent of this rule that competitors should be allowed to use the same data as each other. With this in mind it might it seems only fair that any data that has been posted to the external data board by other competitors should be allowed to be used by anyone.
\n\nBased on the wording of the rules however;
\n\n\"If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.\"
\n\nIt seems that if the data can only be used by the people that actually posted about using it prior to the 13th. Because of that I am retraining my model from scratch so as to not be using the image net weights but for future competitions it might be desirable to change the rules so that in situations like this people may still use the data that other people are using even if they didn't post about it personally before the deadline.
", "votes": 0, "replies": [ { "id": 222730, "author_name": "Tim Joseph", "author_url": "", "post_date": "2017-09-19T19:41:25.827000", "content": "Well noone enforces this rule if you are not a price winner I think. And winning a price this late in the competition seems unlikely, so I would not worry to much!
", "votes": 1, "replies": [] }, { "id": 222739, "author_name": "Tim Anderton", "author_url": "", "post_date": "2017-09-19T20:29:45.903000", "content": "While I agree with you in principle. It just isn't as much fun to work on the problem if the idea of coming out at the top of the leader board causes me distress instead of pleasure.
", "votes": 0, "replies": [] }, { "id": 222875, "author_name": "Artem.Sanakoev", "author_url": "", "post_date": "2017-09-20T12:31:45.563000", "content": "@ Tim, I think you are just nitpicking. For me it seems OK to use data if somebody posted it here one week before the deadline. Why would everyone duplicate the same information here? \nSo I'm sure you shouldn't worry about using Imagenet pretrained weights.
", "votes": 3, "replies": [] }, { "id": 222879, "author_name": "Guillermo Barbadillo", "author_url": "", "post_date": "2017-09-20T12:42:25.653000", "content": "However it is important, because if someone in a good position like you is using external data that is valuable information for all the other participants.
\n\nThat's why the rules are very clear as posted previously
\n\n\n\n\nUse of external data is permitted, provided the data is freely available. If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.
\n
@Artem.Sanakoev: I did not disagree with you at all :)
", "votes": 1, "replies": [] }, { "id": 222979, "author_name": "Artem.Sanakoev", "author_url": "", "post_date": "2017-09-20T19:05:31.483000", "content": "Didn't notice that there are 2 Tims in the thread:)\nI meant @Tim Anderton :)
", "votes": 1, "replies": [] }, { "id": 223188, "author_name": "Tim Joseph", "author_url": "", "post_date": "2017-09-21T11:22:56.040000", "content": "Oh wow, me neither. :D
", "votes": 0, "replies": [] } ] }, { "id": 222661, "author_name": "Ming", "author_url": "", "post_date": "2017-09-19T16:14:49.407000", "content": "ResNet50
", "votes": 0, "replies": [] }, { "id": 222514, "author_name": "SSKH", "author_url": "", "post_date": "2017-09-19T04:55:09.040000", "content": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models
", "votes": 0, "replies": [] }, { "id": 222337, "author_name": "Selim Seferbekov", "author_url": "", "post_date": "2017-09-18T13:53:35.287000", "content": "Pretrained models from https://keras.io/applications/
", "votes": 0, "replies": [] }, { "id": 222332, "author_name": "Vilmar", "author_url": "", "post_date": "2017-09-18T13:45:09.993000", "content": "I might use extra data from this source: https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781
", "votes": 0, "replies": [ { "id": 222333, "author_name": "Artem.Sanakoev", "author_url": "", "post_date": "2017-09-18T13:47:30.603000", "content": "Looks like it's too late. It's less than 1 week till merger deadline
", "votes": 0, "replies": [] }, { "id": 222395, "author_name": "James Requa", "author_url": "", "post_date": "2017-09-18T17:26:12.147000", "content": "I believe the deadline is actually less than 1 week prior to entry deadline, not the merger deadline...so there is still time :)
", "votes": 0, "replies": [] }, { "id": 222401, "author_name": "Guillermo Barbadillo", "author_url": "", "post_date": "2017-09-18T17:38:32.173000", "content": "As far as I understand the last day for publishing external data was 13 September (Entry deadline is 20 September)
\n\nhttps://www.kaggle.com/c/carvana-image-masking-challenge/rules
\n\n\n", "votes": 2, "replies": [] }, { "id": 222411, "author_name": "James Requa", "author_url": "", "post_date": "2017-09-18T18:37:41.677000", "content": "Use of external data is permitted, provided the data is freely available. If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.
\n
OK my bad I thought \"entry deadline\" meant \"submission entry\", but gtk that entry in this context means entry into the competition....If it had just said \"merger deadline\" I think it would have been a lot more clear :)
", "votes": 1, "replies": [] } ] }, { "id": 222255, "author_name": "Konstantin Kolokolov", "author_url": "", "post_date": "2017-09-18T07:21:57.277000", "content": "Pre-trained models from https://github.com/facebook/fb.resnet.torch/tree/master/pretrained and \nhttps://github.com/fchollet/deep-learning-models
", "votes": 0, "replies": [] }, { "id": 222203, "author_name": "Junhong Xu", "author_url": "", "post_date": "2017-09-18T03:11:04.867000", "content": "Pretrained model: VGG-16 with BN, Resnet-34, Resnet-50
", "votes": 0, "replies": [] }, { "id": 221446, "author_name": "Andrés Miguel Torrubia Sáez", "author_url": "", "post_date": "2017-09-15T09:59:28.913000", "content": "Resnet-152 weights for Keras https://gist.github.com/flyyufelix/7e2eafb149f72f4d38dd661882c554a6
", "votes": 0, "replies": [] }, { "id": 220845, "author_name": "Prasanna Hegde", "author_url": "", "post_date": "2017-09-13T08:53:44.147000", "content": "Pre-trained models from tensorflow object detection api https://github.com/tensorflow/models/blob/master/object_detection/g3doc/detection_model_zoo.md
", "votes": 0, "replies": [] }, { "id": 220612, "author_name": "Sean Soleyman", "author_url": "", "post_date": "2017-09-12T17:12:57.310000", "content": "Models posted by Marvin Teichmann: \nhttps://github.com/MarvinTeichmann/KittiSeg\nhttps://github.com/MarvinTeichmann/tensorflow-fcn
\n\nPre-trained model data for these models: \nftp://mi.eng.cam.ac.uk/pub/mttt2/models/
\n\nThese models are TensorFlow implementations and examples of the network architecture described in \"Fully Convolutional Networks for Semantic Segmentation\" by Jonathan Long, Evan Shelhamer, and Trevor Darrell from UC Berkeley: https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf
", "votes": 0, "replies": [] }, { "id": 220509, "author_name": "Avinash Rai", "author_url": "", "post_date": "2017-09-12T11:20:19.183000", "content": "Pre-trained models from Deeplab https://bitbucket.org/deeplab/deeplab-public/
", "votes": 0, "replies": [] }, { "id": 219711, "author_name": "Alex Parinov", "author_url": "", "post_date": "2017-09-09T08:56:05.580000", "content": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models
", "votes": 0, "replies": [] }, { "id": 219470, "author_name": "See--", "author_url": "", "post_date": "2017-09-08T09:56:11.723000", "content": "Pre-trained mobilenet model/weights: https://github.com/fchollet/keras/blob/master/keras/applications/mobilenet.py
", "votes": 0, "replies": [] }, { "id": 218193, "author_name": "XiaokangWang", "author_url": "", "post_date": "2017-09-02T21:39:39.887000", "content": "In the rule it was said external data is permitted. If I google for images of cars online and added them to the training set. Is it allowed?
", "votes": 0, "replies": [ { "id": 218254, "author_name": "Guillermo Barbadillo", "author_url": "", "post_date": "2017-09-03T06:22:46.263000", "content": "In other challenges it was allowed but you had to publish the links to the images and to the masks.
", "votes": 0, "replies": [] } ] }, { "id": 215807, "author_name": "JianWang", "author_url": "", "post_date": "2017-08-23T06:48:21.073000", "content": "I was using the FCN for segmentation however it is relatively slow. Thanks to your idea and I decided to try the Segnet.
", "votes": 0, "replies": [] }, { "id": 207563, "author_name": "Ragnar", "author_url": "", "post_date": "2017-07-26T22:19:11.427000", "content": "Never had a chance to try SegNet and I'm waiting to see what kind of models are going to use, this competition is quite exciting.
\n\nHappy Kaggling!
", "votes": 0, "replies": [] }, { "id": 207531, "author_name": "Marcin Pękalski", "author_url": "", "post_date": "2017-07-26T20:40:20.770000", "content": "\"Creative Commons Attribution-NonCommercial 4.0 International License\"
\n\nCan it be used as it is intended only for non-commercial use?
", "votes": 0, "replies": [ { "id": 207549, "author_name": "Ankasor", "author_url": "", "post_date": "2017-07-26T21:14:20.860000", "content": "By my interpretation, I would assume not: https://wiki.creativecommons.org/wiki/NonCommercial_interpretation
", "votes": 0, "replies": [] } ] }, { "id": 219823, "author_name": "", "author_url": "", "post_date": "2017-09-09T19:27:40.720000", "content": "", "votes": 0, "replies": [] }, { "id": 218816, "author_name": "", "author_url": "", "post_date": "2017-09-06T00:27:15.723000", "content": "", "votes": 0, "replies": [] }, { "id": 218130, "author_name": "", "author_url": "", "post_date": "2017-09-02T13:01:43.540000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "207526": " Competition rules say: \n\n```\nUse of external data is permitted, provided the data is freely available. If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.\n```\n\nHowever, I don't see an official forum thread - so I will start here. :)\n\nI would like to use SegNet: https://github.com/alexgkendall/caffe-segnet\n\n", "222334": "A little bit late, but we have managed to hack GTA-V and generate synthetic data!\nhttps://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781", "222926": "Note to organizers: you should probably FREEZE posting new messages in this thread when the deadline hit, it would avoid any doubts", "218252": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models", "215801": "GTA5 image and semantic label datasets: https://download.visinf.tu-darmstadt.de/data/from_games/", "222396": "Keras pre-trained models https://keras.io/applications/\n\nMight also use this external dataset https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781", "208803": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models", "207528": "Thanks for kicking it off! I've added \"official\" to the title and pinned this thread.", "223431": "Using the crf-cnn pretrained model https://github.com/torrvision/crfasrnn", "222765": "The usual:\n\n- PyTorch: https://github.com/pytorch/vision/tree/master/torchvision/models\n- Standard Keras models: https://keras.io/applications/\n- Various other models: https://github.com/flyyufelix/cnn_finetune\n\nI hope it's not too late! Typically the model posting dateline is by entry deadline or 1 week before the end of competition but it is worded differently here. Anyway, these models have been used in many other previous competitions and have been posted by others earlier.", "222736": "Might use this external data: https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781", "222666": "I just started working on this competition and have built a solution based on using the pretrained models available in the keras applications module. https://keras.io/applications/ I figured that if I could still join the competition that I would also be able to use data that other people are using; if not then what is the point of requiring that we post our data sources? However it is my reading of the rules that I will not be able to use my current solution because I did not personally post that I was using this data a week ago. However it is clearly the intent of this rule that competitors should be allowed to use the same data as each other. With this in mind it might it seems only fair that any data that has been posted to the external data board by other competitors should be allowed to be used by anyone. \n\nBased on the wording of the rules however;\n\n \"If you are using a source of external data, you must post the source to the official external data forum thread no later than one week prior to the entry deadline.\" \n\nIt seems that if the data can only be used by the people that actually posted about using it prior to the 13th. Because of that I am retraining my model from scratch so as to not be using the image net weights but for future competitions it might be desirable to change the rules so that in situations like this people may still use the data that other people are using even if they didn't post about it personally before the deadline.", "222661": "ResNet50", "222514": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models", "222337": "Pretrained models from https://keras.io/applications/", "222332": "I might use extra data from this source: https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781", "222255": "Pre-trained models from https://github.com/facebook/fb.resnet.torch/tree/master/pretrained and \nhttps://github.com/fchollet/deep-learning-models", "222203": "Pretrained model: VGG-16 with BN, Resnet-34, Resnet-50", "221446": "Resnet-152 weights for Keras https://gist.github.com/flyyufelix/7e2eafb149f72f4d38dd661882c554a6", "220845": "Pre-trained models from tensorflow object detection api https://github.com/tensorflow/models/blob/master/object_detection/g3doc/detection_model_zoo.md", "220612": "Models posted by Marvin Teichmann: \nhttps://github.com/MarvinTeichmann/KittiSeg\nhttps://github.com/MarvinTeichmann/tensorflow-fcn\n\nPre-trained model data for these models: \nftp://mi.eng.cam.ac.uk/pub/mttt2/models/\n\nThese models are TensorFlow implementations and examples of the network architecture described in \"Fully Convolutional Networks for Semantic Segmentation\" by Jonathan Long, Evan Shelhamer, and Trevor Darrell from UC Berkeley: https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf", "220509": "Pre-trained models from Deeplab https://bitbucket.org/deeplab/deeplab-public/", "219711": "Pre-trained models from torchvision: https://github.com/pytorch/vision/tree/master/torchvision/models", "219470": "Pre-trained mobilenet model/weights: https://github.com/fchollet/keras/blob/master/keras/applications/mobilenet.py", "218193": "In the rule it was said external data is permitted. If I google for images of cars online and added them to the training set. Is it allowed? ", "215807": "I was using the FCN for segmentation however it is relatively slow. Thanks to your idea and I decided to try the Segnet.", "207563": "Never had a chance to try SegNet and I'm waiting to see what kind of models are going to use, this competition is quite exciting.\n\nHappy Kaggling!", "207531": "\"Creative Commons Attribution-NonCommercial 4.0 International License\"\n\nCan it be used as it is intended only for non-commercial use?", "219823": "", "218816": "", "218130": "" } }