{ "id": 236155, "title": "External Data Disclosure Thread", "url": "/competitions/hotel-id-2021-fgvc8/discussion/236155", "author_name": "", "post_date": "2021-05-03T05:19:46.996827900Z", "votes": 5, "comment_count": 5, "views": 0, "content": "

Here is the list of external datasets that I might use:

\n\n

Good luck!

", "messages": [ { "id": "1291510", "postDate": "05/03/2021 05:19:46", "content": "

Here is the list of external datasets that I might use:

\n\n

Good luck!

", "rawMarkdown": "Here is the list of external datasets that I might use:\n\n* timm pretrained models https://github.com/rwightman/pytorch-image-models\n* mmdetection pretrained models https://github.com/open-mmlab/mmdetection\n* LoFTR pre-trained models https://github.com/zju3dv/LoFTR\n* SuperGLUE/SuperPoint pretrained models https://github.com/magicleap/SuperGluePretrainedNetwork\n* DELG pre-trained models https://github.com/tensorflow/models/blob/master/research/delf/delf/python/delg/DELG_INSTRUCTIONS.md\n* Places365 dataset https://github.com/CSAILVision/places365\n* OpenImages dataset https://storage.googleapis.com/openimages/web/index.html\n* Hotels50k dataset https://github.com/GWUvision/Hotels-50K\n\nGood luck!", "votes": null }, { "id": "1297140", "postDate": "05/07/2021 18:48:58", "content": "

@confirm thanks for sharing the links.

\n

@maggiemd I'm a bit confused about the comp rules.

\n

A3 says:

\n
\n

The general rule is that participants should only use the provided training images for training models to classify the test images. We do not want participants crawling the web in search of additional data.

\n
\n

But B7C contains the general external data clause.

\n

So, are we allowed to use data such as Hotels50k dataset https://github.com/GWUvision/Hotels-50K?

", "rawMarkdown": "confirm thanks for sharing the links.\n\n@maggiemd I'm a bit confused about the comp rules.\n\nA3 says:\n> The general rule is that participants should only use the provided training images for training models to classify the test images. We do not want participants crawling the web in search of additional data. \n\nBut B7C contains the general **external data** clause.\n\nSo, are we allowed to use data such as Hotels50k dataset https://github.com/GWUvision/Hotels-50K?", "votes": null }, { "id": "1300784", "postDate": "05/10/2021 17:54:05", "content": "

Hi, thank you for sharing these very useful resources! I don't know if you can solve my problem but I can try to explain it:
\nI used a simple classifier with efficientnet b2 on hotel id classes and I obtained the score on LB. When I try very advanced techniques I can't improve my mAP. 1) I tried to group rooms and bathrooms and use different classifiers on it
\n2) I tried to group similar perspective images inside every hotels and use it with a classifier
\n3) I tried arc margin loss (the worst result) in cases 1) and 2)
\n4) I tried to clean the dataset, discovering outliers using superglue and object detection

\n

I don't know what I missed, but it's very frustrating to see that a simple softmax without anything is my best result… I don't want to know the solution (I really don't care about LB scoring but the problem itself) but if you have some tips it could be very appreciated :) Thanks!

", "rawMarkdown": "Hi, thank you for sharing these very useful resources! I don't know if you can solve my problem but I can try to explain it:\nI used a simple classifier with efficientnet b2 on hotel id classes and I obtained the score on LB. When I try very advanced techniques I can't improve my mAP. 1) I tried to group rooms and bathrooms and use different classifiers on it\n2) I tried to group similar perspective images inside every hotels and use it with a classifier\n3) I tried arc margin loss (the worst result) in cases 1) and 2)\n4) I tried to clean the dataset, discovering outliers using superglue and object detection\n\nI don't know what I missed, but it's very frustrating to see that a simple softmax without anything is my best result... I don't want to know the solution (I really don't care about LB scoring but the problem itself) but if you have some tips it could be very appreciated :) Thanks!", "votes": null }, { "id": "1302833", "postDate": "05/11/2021 18:35:02", "content": "

I'm probably using:

\n

Pretrained models from:

\n", "rawMarkdown": "I'm probably using:\n\nPretrained models from:\n\n- https://pytorch.org/vision/stable/models.html (via fastai)\n- https://github.com/lukemelas/EfficientNet-PyTorch", "votes": null }, { "id": "1324439", "postDate": "05/27/2021 00:23:59", "content": "

Congratulate !

", "rawMarkdown": "Congratulate !", "votes": null }, { "id": "1324857", "postDate": "05/27/2021 09:34:35", "content": "

Congratulate !

", "rawMarkdown": "Congratulate !", "votes": null } ], "comments": [ { "id": 1297140, "author_name": "joatom", "author_url": "", "post_date": "05/07/2021 18:48:58", "content": "

@confirm thanks for sharing the links.

\n

@maggiemd I'm a bit confused about the comp rules.

\n

A3 says:

\n
\n

The general rule is that participants should only use the provided training images for training models to classify the test images. We do not want participants crawling the web in search of additional data.

\n
\n

But B7C contains the general external data clause.

\n

So, are we allowed to use data such as Hotels50k dataset https://github.com/GWUvision/Hotels-50K?

", "votes": null, "replies": [] }, { "id": 1300784, "author_name": "alenic", "author_url": "", "post_date": "05/10/2021 17:54:05", "content": "

Hi, thank you for sharing these very useful resources! I don't know if you can solve my problem but I can try to explain it:
\nI used a simple classifier with efficientnet b2 on hotel id classes and I obtained the score on LB. When I try very advanced techniques I can't improve my mAP. 1) I tried to group rooms and bathrooms and use different classifiers on it
\n2) I tried to group similar perspective images inside every hotels and use it with a classifier
\n3) I tried arc margin loss (the worst result) in cases 1) and 2)
\n4) I tried to clean the dataset, discovering outliers using superglue and object detection

\n

I don't know what I missed, but it's very frustrating to see that a simple softmax without anything is my best result… I don't want to know the solution (I really don't care about LB scoring but the problem itself) but if you have some tips it could be very appreciated :) Thanks!

", "votes": null, "replies": [] }, { "id": 1302833, "author_name": "joatom", "author_url": "", "post_date": "05/11/2021 18:35:02", "content": "

I'm probably using:

\n

Pretrained models from:

\n", "votes": null, "replies": [] }, { "id": 1324439, "author_name": "arthasmenethil", "author_url": "", "post_date": "05/27/2021 00:23:59", "content": "

Congratulate !

", "votes": null, "replies": [] }, { "id": 1324857, "author_name": "sicelice", "author_url": "", "post_date": "05/27/2021 09:34:35", "content": "

Congratulate !

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1291510": "Here is the list of external datasets that I might use:\n\n* timm pretrained models https://github.com/rwightman/pytorch-image-models\n* mmdetection pretrained models https://github.com/open-mmlab/mmdetection\n* LoFTR pre-trained models https://github.com/zju3dv/LoFTR\n* SuperGLUE/SuperPoint pretrained models https://github.com/magicleap/SuperGluePretrainedNetwork\n* DELG pre-trained models https://github.com/tensorflow/models/blob/master/research/delf/delf/python/delg/DELG_INSTRUCTIONS.md\n* Places365 dataset https://github.com/CSAILVision/places365\n* OpenImages dataset https://storage.googleapis.com/openimages/web/index.html\n* Hotels50k dataset https://github.com/GWUvision/Hotels-50K\n\nGood luck!", "1297140": "confirm thanks for sharing the links.\n\n@maggiemd I'm a bit confused about the comp rules.\n\nA3 says:\n> The general rule is that participants should only use the provided training images for training models to classify the test images. We do not want participants crawling the web in search of additional data. \n\nBut B7C contains the general **external data** clause.\n\nSo, are we allowed to use data such as Hotels50k dataset https://github.com/GWUvision/Hotels-50K?", "1300784": "Hi, thank you for sharing these very useful resources! I don't know if you can solve my problem but I can try to explain it:\nI used a simple classifier with efficientnet b2 on hotel id classes and I obtained the score on LB. When I try very advanced techniques I can't improve my mAP. 1) I tried to group rooms and bathrooms and use different classifiers on it\n2) I tried to group similar perspective images inside every hotels and use it with a classifier\n3) I tried arc margin loss (the worst result) in cases 1) and 2)\n4) I tried to clean the dataset, discovering outliers using superglue and object detection\n\nI don't know what I missed, but it's very frustrating to see that a simple softmax without anything is my best result... I don't want to know the solution (I really don't care about LB scoring but the problem itself) but if you have some tips it could be very appreciated :) Thanks!", "1302833": "I'm probably using:\n\nPretrained models from:\n\n- https://pytorch.org/vision/stable/models.html (via fastai)\n- https://github.com/lukemelas/EfficientNet-PyTorch", "1324439": "Congratulate !", "1324857": "Congratulate !" }, "source": "meta" }