{ "id": 394959, "title": "Adding negative examples to the training dataset", "url": "/competitions/asl-signs/discussion/394959", "author_name": "", "post_date": "2023-03-15T11:36:18.445729600Z", "votes": null, "comment_count": 4, "views": 0, "content": "

Has anyone tried adding some non-sign MediaPipe data to the dataset as a separate \"negative\" class? I've heard that adding 10-15% negative samples could improve the model performance.

", "messages": [ { "id": "2182917", "postDate": "03/15/2023 11:36:18", "content": "

Has anyone tried adding some non-sign MediaPipe data to the dataset as a separate \"negative\" class? I've heard that adding 10-15% negative samples could improve the model performance.

", "rawMarkdown": "Has anyone tried adding some non-sign MediaPipe data to the dataset as a separate \"negative\" class? I've heard that adding 10-15% negative samples could improve the model performance.", "votes": null }, { "id": "2182953", "postDate": "03/15/2023 11:58:08", "content": "

May I ask how to get the non-sign MediaPipe data

", "rawMarkdown": "May I ask how to get the non-sign MediaPipe data", "votes": null }, { "id": "2182963", "postDate": "03/15/2023 12:07:02", "content": "

MediaPipe docs mention Youtube-8M challenge and this repo for landmark extraction, I'm not sure if there's a processed dataset though, because it's set up to extract different kind of landmarks (RGB features instead of pose features).

", "rawMarkdown": "MediaPipe docs mention [Youtube-8M challenge](https://google.github.io/mediapipe/solutions/youtube_8m.html#extracting-video-features-for-youtube-8m-challenge) and [this repo](https://github.com/google/youtube-8m/tree/master/feature_extractor) for landmark extraction, I'm not sure if there's a processed dataset though, because it's set up to extract different kind of landmarks (RGB features instead of pose features).", "votes": null }, { "id": "2182973", "postDate": "03/15/2023 12:11:27", "content": "

flip the parquet file (mirror left right) and treated them as unlabelled data
\n(i think some hand sign are symmetrical and the left-right hand can interchange)

\n

but i am more into self-supervised learning with transformer
\ni am going to use generative prediction (predict the next frame skeleton landmark) and let the model to watch millions of youtube ASL videos

", "rawMarkdown": "flip the parquet file (mirror left right) and treated them as unlabelled data\n(i think some hand sign are symmetrical and the left-right hand can interchange)\n\nbut i am more into self-supervised learning with transformer\ni am going to use generative prediction (predict the next frame skeleton landmark) and let the model to watch millions of youtube ASL videos", "votes": null }, { "id": "2184980", "postDate": "03/16/2023 18:10:10", "content": "

just realise that reverse time can also be used as unlabelled data

", "rawMarkdown": "just realise that reverse time can also be used as unlabelled data", "votes": null } ], "comments": [ { "id": 2182953, "author_name": "sayoulala", "author_url": "", "post_date": "03/15/2023 11:58:08", "content": "

May I ask how to get the non-sign MediaPipe data

", "votes": null, "replies": [ { "id": 2182963, "author_name": "vbogach", "author_url": "", "post_date": "03/15/2023 12:07:02", "content": "

MediaPipe docs mention Youtube-8M challenge and this repo for landmark extraction, I'm not sure if there's a processed dataset though, because it's set up to extract different kind of landmarks (RGB features instead of pose features).

", "votes": null, "replies": [] } ] }, { "id": 2182973, "author_name": "hengck23", "author_url": "", "post_date": "03/15/2023 12:11:27", "content": "

flip the parquet file (mirror left right) and treated them as unlabelled data
\n(i think some hand sign are symmetrical and the left-right hand can interchange)

\n

but i am more into self-supervised learning with transformer
\ni am going to use generative prediction (predict the next frame skeleton landmark) and let the model to watch millions of youtube ASL videos

", "votes": null, "replies": [ { "id": 2184980, "author_name": "hengck23", "author_url": "", "post_date": "03/16/2023 18:10:10", "content": "

just realise that reverse time can also be used as unlabelled data

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2182917": "Has anyone tried adding some non-sign MediaPipe data to the dataset as a separate \"negative\" class? I've heard that adding 10-15% negative samples could improve the model performance.", "2182953": "May I ask how to get the non-sign MediaPipe data", "2182963": "MediaPipe docs mention [Youtube-8M challenge](https://google.github.io/mediapipe/solutions/youtube_8m.html#extracting-video-features-for-youtube-8m-challenge) and [this repo](https://github.com/google/youtube-8m/tree/master/feature_extractor) for landmark extraction, I'm not sure if there's a processed dataset though, because it's set up to extract different kind of landmarks (RGB features instead of pose features).", "2182973": "flip the parquet file (mirror left right) and treated them as unlabelled data\n(i think some hand sign are symmetrical and the left-right hand can interchange)\n\nbut i am more into self-supervised learning with transformer\ni am going to use generative prediction (predict the next frame skeleton landmark) and let the model to watch millions of youtube ASL videos", "2184980": "just realise that reverse time can also be used as unlabelled data" }, "source": "meta" }