![\"https://i.ibb.co/mXs1xfB/Selection-999-1698.png\"](\"https://i.ibb.co/mXs1xfB/Selection-999-1698.png\")
I tried my tflite model on WLASL dataset.
\ni apply mediapose on a few videos and it is wokring well
\n(WLASL has 200 out of 250 kaggle labels)
This open up the use of external dataset
\n(NOTE !!!! you will have to verify the license of of the external set yourself. it is not clear if publice dataset like WLASL can be used)
code to convert video to pose data
\nhttps://www.kaggle.com/code/hengck23/generate-your-own-pose-data-from-video
I tried my tflite model on WLASL dataset.
\ni apply mediapose on a few videos and it is wokring well
\n(WLASL has 200 out of 250 kaggle labels)
This open up the use of external dataset
\n(NOTE !!!! you will have to verify the license of of the external set yourself. it is not clear if publice dataset like WLASL can be used)
code to convert video to pose data
\nhttps://www.kaggle.com/code/hengck23/generate-your-own-pose-data-from-video
https://github.com/dxli94/WLASL/blob/master/README.md
\nhttps://www.kaggle.com/datasets/risangbaskoro/wlasl-processed
Thank you for sharing! I tried to find some externel dataset, but I found that almost every existing sign language dataset does not allow commercial use. I personally think we cannot use those because delivering Competition Sponsor the final model's code without any license issue seems pretty important in this competition. And It would be great if the host could provide clarification on whether we are allowed to use a specific dataset like WLASL. we can find a list of some known sign language datasets here http://facundoq.github.io/guides/sign_language_datasets/slr
", "rawMarkdown": "Thank you for sharing! I tried to find some externel dataset, but I found that almost every existing sign language dataset does not allow commercial use. I personally think we cannot use those because delivering Competition Sponsor the final model's code without any license issue seems pretty important in this competition. And It would be great if the host could provide clarification on whether we are allowed to use a specific dataset like WLASL. we can find a list of some known sign language datasets here http://facundoq.github.io/guides/sign_language_datasets/slr", "votes": null }, { "id": "2210275", "postDate": "04/05/2023 09:01:44", "content": "This will depends on the organizer.
\nIs PopSign a commerical application?
\nThere are some kaggle competition that allows non-commerical public dataset becuase the organizer or application is not commerical use.
Yes. it depends on host's decision whether or not to allow non-commercial dataset or model. But If the host does not clarify anything, I personally won't take a risk in this competition.
", "rawMarkdown": "Yes. it depends on host's decision whether or not to allow non-commercial dataset or model. But If the host does not clarify anything, I personally won't take a risk in this competition.", "votes": null }, { "id": "2210314", "postDate": "04/05/2023 09:38:48", "content": "with a webcam, you can be the signer can collect the data yourself.
\nyou can install popsign in your mobile and practice the handsign sign , etc
\n(or you can have the mobile camera and webcam capturing you at the same time)
with 847 teams, if each team contribute 5 videos or pose, we will have additional 3k data per class
\n(with generative AI like stable diffusion, it is easy to turn one video to multiple video)
if there is expert on android emulator linux, please advise on how to setup popsign on PC to collect data
", "rawMarkdown": "if there is expert on android emulator linux, please advise on how to setup popsign on PC to collect data", "votes": null }, { "id": "2210320", "postDate": "04/05/2023 09:45:58", "content": "enb01more-run200-tx030d-aug5-seed.fp16.tflite (LB: 0.76+)
\nnum of WLASL videos tested:\n1512\n\ntop-5 (use 250 class):\n[0.61640212 0.71560847 0.75793651 0.77579365 0.78902116]\n\ntop-5 (use 200 class):\n[0.6521164 0.74140212 0.77380952 0.79695767 0.81415344]\nexpected LB improvement if adding external data: +0.01
", "rawMarkdown": "enb01more-run200-tx030d-aug5-seed.fp16.tflite (LB: 0.76+)\n\n\n```\nnum of WLASL videos tested:\n1512\n\ntop-5 (use 250 class):\n[0.61640212 0.71560847 0.75793651 0.77579365 0.78902116]\n\ntop-5 (use 200 class):\n[0.6521164 0.74140212 0.77380952 0.79695767 0.81415344]\n\n```\n\nexpected LB improvement if adding external data: +0.01", "votes": null }, { "id": "2210335", "postDate": "04/05/2023 10:05:17", "content": "![\"https://i.ibb.co/b1kQs9k/Selection-999-1699.png\"](\"https://i.ibb.co/b1kQs9k/Selection-999-1699.png\")
\nhttps://arxiv.org/pdf/2303.12080v1.pdf
\nNatural Language-Assisted Sign Language Recognition
![\"https://i.ibb.co/CsNKw3j/Selection-999-1701.png\"](\"https://i.ibb.co/CsNKw3j/Selection-999-1701.png\")
\nBEST: BERT Pre-Training for Sign Language Recognition with Coupling Tokenization
WLASL seems to be released under C-UDA.
\nhttps://news.microsoft.com/wp-content/uploads/prod/sites/560/2019/12/C-UDA_README.pdf
I think it can be used to train commercial models unless we don't redistribute the dataset, but I won't use it unless the host clarifies the usage. (remember deepfake)
", "rawMarkdown": "WLASL seems to be released under C-UDA.\nhttps://news.microsoft.com/wp-content/uploads/prod/sites/560/2019/12/C-UDA_README.pdf\n\nI think it can be used to train commercial models unless we don't redistribute the dataset, but I won't use it unless the host clarifies the usage. (remember deepfake)", "votes": null }, { "id": "2210905", "postDate": "04/05/2023 17:27:51", "content": "Thank you for the question! External Data needs to allow you to license the winning solution, i.e. being used in commercial purpose. Thus, you may use any external data as long as it is public and commercially available. Thanks so much!
", "rawMarkdown": "Thank you for the question! External Data needs to allow you to license the winning solution, i.e. being used in commercial purpose. Thus, you may use any external data as long as it is public and commercially available. Thanks so much!", "votes": null }, { "id": "2210916", "postDate": "04/05/2023 17:40:42", "content": "\n\nAll the WLASL data is intended for academic and computational use only. No commercial usage is allowed. We highly respect copyright and privacy. If you find WLASL violates your rights, please contact us.
\n
\n (from https://github.com/dxli94/WLASL)
@ashleychow That means we cannot use WLASL, right ?
", "rawMarkdown": "> All the WLASL data is intended for academic and computational use only. No commercial usage is allowed. We highly respect copyright and privacy. If you find WLASL violates your rights, please contact us.\n(from https://github.com/dxli94/WLASL)\n\n@ashleychow That means we cannot use WLASL, right ?", "votes": null }, { "id": "2210957", "postDate": "04/05/2023 18:23:21", "content": "Looks like we cannot use this dataset. As it is not commercially available. Please clarify?
", "rawMarkdown": "Looks like we cannot use this dataset. As it is not commercially available. Please clarify?", "votes": null }, { "id": "2211376", "postDate": "04/06/2023 02:00:49", "content": "while public data may not be used to develop and submit as solution, public data is definitely useful for algorithm debug and determine the robustness/weakness of your model.
\nafter you know the weakness of your model, you would need to solve it and submit without the use of external data.
\nbesides WLASl, i also try MSASL
\nhttps://www.microsoft.com/en-us/research/project/ms-asl/
\nhttps://github.com/iamgarcia/msasl-video-downloader
i observed the following weakness:
\nnum of samples (i only tested 2k out of 3k of overlap signs)\n\npredict (2256, 250)\ntruth (2256,)\n\nall class 250\n[0.5589539 0.64007092 0.68085106 0.70833333 0.7251773 ]\n\nsub class 154\n[0.60549645 0.69060284 0.72163121 0.74778369 0.76019504]\nhttps://github.com/dxli94/WLASL/blob/master/README.md
\nhttps://www.kaggle.com/datasets/risangbaskoro/wlasl-processed
Thank you for sharing! I tried to find some externel dataset, but I found that almost every existing sign language dataset does not allow commercial use. I personally think we cannot use those because delivering Competition Sponsor the final model's code without any license issue seems pretty important in this competition. And It would be great if the host could provide clarification on whether we are allowed to use a specific dataset like WLASL. we can find a list of some known sign language datasets here http://facundoq.github.io/guides/sign_language_datasets/slr
", "votes": null, "replies": [ { "id": 2210275, "author_name": "hengck23", "author_url": "", "post_date": "04/05/2023 09:01:44", "content": "This will depends on the organizer.
\nIs PopSign a commerical application?
\nThere are some kaggle competition that allows non-commerical public dataset becuase the organizer or application is not commerical use.
Yes. it depends on host's decision whether or not to allow non-commercial dataset or model. But If the host does not clarify anything, I personally won't take a risk in this competition.
", "votes": null, "replies": [ { "id": 2210314, "author_name": "hengck23", "author_url": "", "post_date": "04/05/2023 09:38:48", "content": "with a webcam, you can be the signer can collect the data yourself.
\nyou can install popsign in your mobile and practice the handsign sign , etc
\n(or you can have the mobile camera and webcam capturing you at the same time)
with 847 teams, if each team contribute 5 videos or pose, we will have additional 3k data per class
\n(with generative AI like stable diffusion, it is easy to turn one video to multiple video)
if there is expert on android emulator linux, please advise on how to setup popsign on PC to collect data
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2210905, "author_name": "ashleychow", "author_url": "", "post_date": "04/05/2023 17:27:51", "content": "Thank you for the question! External Data needs to allow you to license the winning solution, i.e. being used in commercial purpose. Thus, you may use any external data as long as it is public and commercially available. Thanks so much!
", "votes": null, "replies": [ { "id": 2210916, "author_name": "theoviel", "author_url": "", "post_date": "04/05/2023 17:40:42", "content": "\n\nAll the WLASL data is intended for academic and computational use only. No commercial usage is allowed. We highly respect copyright and privacy. If you find WLASL violates your rights, please contact us.
\n
\n (from https://github.com/dxli94/WLASL)
@ashleychow That means we cannot use WLASL, right ?
", "votes": null, "replies": [] } ] } ] }, { "id": 2210320, "author_name": "hengck23", "author_url": "", "post_date": "04/05/2023 09:45:58", "content": "enb01more-run200-tx030d-aug5-seed.fp16.tflite (LB: 0.76+)
\nnum of WLASL videos tested:\n1512\n\ntop-5 (use 250 class):\n[0.61640212 0.71560847 0.75793651 0.77579365 0.78902116]\n\ntop-5 (use 200 class):\n[0.6521164 0.74140212 0.77380952 0.79695767 0.81415344]\nexpected LB improvement if adding external data: +0.01
", "votes": null, "replies": [ { "id": 2210335, "author_name": "hengck23", "author_url": "", "post_date": "04/05/2023 10:05:17", "content": "
\nhttps://arxiv.org/pdf/2303.12080v1.pdf
\nNatural Language-Assisted Sign Language Recognition
\nBEST: BERT Pre-Training for Sign Language Recognition with Coupling Tokenization
while public data may not be used to develop and submit as solution, public data is definitely useful for algorithm debug and determine the robustness/weakness of your model.
\nafter you know the weakness of your model, you would need to solve it and submit without the use of external data.
\nbesides WLASl, i also try MSASL
\nhttps://www.microsoft.com/en-us/research/project/ms-asl/
\nhttps://github.com/iamgarcia/msasl-video-downloader
i observed the following weakness:
\nnum of samples (i only tested 2k out of 3k of overlap signs)\n\npredict (2256, 250)\ntruth (2256,)\n\nall class 250\n[0.5589539 0.64007092 0.68085106 0.70833333 0.7251773 ]\n\nsub class 154\n[0.60549645 0.69060284 0.72163121 0.74778369 0.76019504]\nWLASL seems to be released under C-UDA.
\nhttps://news.microsoft.com/wp-content/uploads/prod/sites/560/2019/12/C-UDA_README.pdf
I think it can be used to train commercial models unless we don't redistribute the dataset, but I won't use it unless the host clarifies the usage. (remember deepfake)
", "votes": null, "replies": [] }, { "id": 2210957, "author_name": "mannyjain", "author_url": "", "post_date": "04/05/2023 18:23:21", "content": "Looks like we cannot use this dataset. As it is not commercially available. Please clarify?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2210137": "I tried my tflite model on WLASL dataset.\ni apply mediapose on a few videos and it is wokring well\n(WLASL has 200 out of 250 kaggle labels)\n\nThis open up the use of external dataset \n(NOTE !!!! you will have to verify the license of of the external set yourself. it is not clear if publice dataset like WLASL can be used)\n\ncode to convert video to pose data\nhttps://www.kaggle.com/code/hengck23/generate-your-own-pose-data-from-video\n\n", "2210241": "https://github.com/dxli94/WLASL/blob/master/README.md\nhttps://www.kaggle.com/datasets/risangbaskoro/wlasl-processed", "2210263": "Thank you for sharing! I tried to find some externel dataset, but I found that almost every existing sign language dataset does not allow commercial use. I personally think we cannot use those because delivering Competition Sponsor the final model's code without any license issue seems pretty important in this competition. And It would be great if the host could provide clarification on whether we are allowed to use a specific dataset like WLASL. we can find a list of some known sign language datasets here http://facundoq.github.io/guides/sign_language_datasets/slr", "2210275": "This will depends on the organizer.\n\nIs PopSign a commerical application?\nThere are some kaggle competition that allows non-commerical public dataset becuase the organizer or application is not commerical use.", "2210296": "Yes. it depends on host's decision whether or not to allow non-commercial dataset or model. But If the host does not clarify anything, I personally won't take a risk in this competition.", "2210314": "with a webcam, you can be the signer can collect the data yourself.\nyou can install popsign in your mobile and practice the handsign sign , etc \n(or you can have the mobile camera and webcam capturing you at the same time)\n\nwith 847 teams, if each team contribute 5 videos or pose, we will have additional 3k data per class\n(with generative AI like stable diffusion, it is easy to turn one video to multiple video)", "2210318": "if there is expert on android emulator linux, please advise on how to setup popsign on PC to collect data", "2210320": "enb01more-run200-tx030d-aug5-seed.fp16.tflite (LB: 0.76+)\n\n\n```\nnum of WLASL videos tested:\n1512\n\ntop-5 (use 250 class):\n[0.61640212 0.71560847 0.75793651 0.77579365 0.78902116]\n\ntop-5 (use 200 class):\n[0.6521164 0.74140212 0.77380952 0.79695767 0.81415344]\n\n```\n\nexpected LB improvement if adding external data: +0.01", "2210335": "\nhttps://arxiv.org/pdf/2303.12080v1.pdf\nNatural Language-Assisted Sign Language Recognition\n\n\nBEST: BERT Pre-Training for Sign Language Recognition with Coupling Tokenization", "2210509": "WLASL seems to be released under C-UDA.\nhttps://news.microsoft.com/wp-content/uploads/prod/sites/560/2019/12/C-UDA_README.pdf\n\nI think it can be used to train commercial models unless we don't redistribute the dataset, but I won't use it unless the host clarifies the usage. (remember deepfake)", "2210905": "Thank you for the question! External Data needs to allow you to license the winning solution, i.e. being used in commercial purpose. Thus, you may use any external data as long as it is public and commercially available. Thanks so much!", "2210916": "> All the WLASL data is intended for academic and computational use only. No commercial usage is allowed. We highly respect copyright and privacy. If you find WLASL violates your rights, please contact us.\n(from https://github.com/dxli94/WLASL)\n\n@ashleychow That means we cannot use WLASL, right ?", "2210957": "Looks like we cannot use this dataset. As it is not commercially available. Please clarify?", "2211376": "while public data **may not be used to develop and submit as solution**, public data is definitely useful for algorithm debug and determine the robustness/weakness of your model.\n\nafter you know the weakness of your model, you would need to solve it and submit **without the use of external data.**\n\n---\n\nbesides WLASl, i also try MSASL\nhttps://www.microsoft.com/en-us/research/project/ms-asl/\nhttps://github.com/iamgarcia/msasl-video-downloader\n\ni observed the following weakness:\n1. kaggle dataset mostly has only one hand visible. when external data shows both hand, prediction is wrong.\n2. when external data shows face that is non-frontal, prediction is wrong.\n\n```\nnum of samples (i only tested 2k out of 3k of overlap signs)\n\npredict (2256, 250)\ntruth (2256,)\n\nall class 250\n[0.5589539 0.64007092 0.68085106 0.70833333 0.7251773 ]\n\nsub class 154\n[0.60549645 0.69060284 0.72163121 0.74778369 0.76019504]\n\n```" }, "source": "meta" }