{ "id": 406300, "title": "12th place solution - MLP Based (Structured Keypoint Pooling network)", "url": "/competitions/asl-signs/writeups/donjyarahoi-12th-place-solution-mlp-based-structur", "author_name": "", "post_date": "2023-05-02T00:17:10.847Z", "votes": 25, "comment_count": 2, "views": 0, "content": "

Thanks to the hosts and the Kaggle team for hosting such an amazing competition.\\
\nBelow is my brife solution.

\n

Model

\n\n

Data

\n\n

Augmentation

\n\n

Loss

\n\n

What Didn't Work

\n\n

Especially thanks to these kernels

\n", "messages": [ { "id": "2241887", "postDate": "05/02/2023 00:04:09", "content": "

Thanks to the hosts and the Kaggle team for hosting such an amazing competition.\\
\nBelow is my brife solution.

\n

Model

\n\n

Data

\n\n

Augmentation

\n\n

Loss

\n\n

What Didn't Work

\n\n

Especially thanks to these kernels

\n", "rawMarkdown": "Thanks to the hosts and the Kaggle team for hosting such an amazing competition.\\\nBelow is my brife solution.\n\n## Model\n\n* MLP based. I used a modified model of [the Structured Keypoint Pooling network](https://arxiv.org/abs/2303.15270).\n* Input joint and bone separately. Bidirectional lateral Connection. Concat after 2 mlp blocks. \n * Refered to [Two-Stream Network for Sign Language Recognition and Translation](https://arxiv.org/abs/2211.01367)\n\n## Data\n* Only original data\n* Lip + Pose (upper body) + Left hand + Right hand\n* Resize to 64 frames, if long.\n* Feature: xy (joint or bone), motion of 1&2frame.\n* So, input shape is [64(frame), (107(joint) + 142(bone)) x 2 (xy) x 3 (+1&2motion)]\n* Exclude data with an low estimation probability on the training data (1%). [CV: -0.006, publicLB +0.01]\n\n## Augmentation\n* Flip [CV: +0.02]\n* Scale [CV: + 0.008]\n* Rotate3d all data [CV: + 0.003]\n * Better than Rotate2d by CV 0.002\n* Rotate3d hands individually [CV: +0.002]\n* Resize frame (only shorten the frame) [CV: + 0.005]\n* Cut the first or last frame [CV: +0.001]\n* Swap hand or lip with same sign. [Hand CV: +0.006, Lip CV: + 0.002]\n\n## Loss\n* CrossEntropy (label smoothing 0.4)\n \n## What Didn't Work\n* Word based-label smoothing\n* Manifold mixup\n* Shift DA\n* Adding eye\n* Pseudo labels\n\n## Especially thanks to these kernels\n* [[LB 0.67] one pytorch transformer solution](https://www.kaggle.com/code/hengck23/lb-0-67-one-pytorch-transformer-solution) [@hengck23](https://www.kaggle.com/hengck23)\n* [GISLR [LB 0.63]: On the Shoulders](https://www.kaggle.com/code/roberthatch/gislr-lb-0-63-on-the-shoulders) [@roberthatch](https://www.kaggle.com/roberthatch)\n* [Animated Data Visualization](https://www.kaggle.com/code/danielpeshkov/animated-data-visualization) [@danielpeshkov](https://www.kaggle.com/danielpeshkov)", "votes": null }, { "id": "2241984", "postDate": "05/02/2023 01:29:18", "content": "

Hearty congratulations for the result and kudos to you for the approach note as well @donjyarahoi

", "rawMarkdown": "Hearty congratulations for the result and kudos to you for the approach note as well @donjyarahoi", "votes": null }, { "id": "2242850", "postDate": "05/02/2023 14:38:41", "content": "

@donjyarahoi congratulations with great result!💥 Thanks for sharing - it helps to improve skills!🤝🙂

", "rawMarkdown": "donjyarahoi congratulations with great result!💥 Thanks for sharing - it helps to improve skills!🤝🙂", "votes": null } ], "comments": [ { "id": 2241984, "author_name": "ravi20076", "author_url": "", "post_date": "05/02/2023 01:29:18", "content": "

Hearty congratulations for the result and kudos to you for the approach note as well @donjyarahoi

", "votes": null, "replies": [] }, { "id": 2242850, "author_name": "ivanisaev", "author_url": "", "post_date": "05/02/2023 14:38:41", "content": "

@donjyarahoi congratulations with great result!💥 Thanks for sharing - it helps to improve skills!🤝🙂

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2241887": "Thanks to the hosts and the Kaggle team for hosting such an amazing competition.\\\nBelow is my brife solution.\n\n## Model\n\n* MLP based. I used a modified model of [the Structured Keypoint Pooling network](https://arxiv.org/abs/2303.15270).\n* Input joint and bone separately. Bidirectional lateral Connection. Concat after 2 mlp blocks. \n * Refered to [Two-Stream Network for Sign Language Recognition and Translation](https://arxiv.org/abs/2211.01367)\n\n## Data\n* Only original data\n* Lip + Pose (upper body) + Left hand + Right hand\n* Resize to 64 frames, if long.\n* Feature: xy (joint or bone), motion of 1&2frame.\n* So, input shape is [64(frame), (107(joint) + 142(bone)) x 2 (xy) x 3 (+1&2motion)]\n* Exclude data with an low estimation probability on the training data (1%). [CV: -0.006, publicLB +0.01]\n\n## Augmentation\n* Flip [CV: +0.02]\n* Scale [CV: + 0.008]\n* Rotate3d all data [CV: + 0.003]\n * Better than Rotate2d by CV 0.002\n* Rotate3d hands individually [CV: +0.002]\n* Resize frame (only shorten the frame) [CV: + 0.005]\n* Cut the first or last frame [CV: +0.001]\n* Swap hand or lip with same sign. [Hand CV: +0.006, Lip CV: + 0.002]\n\n## Loss\n* CrossEntropy (label smoothing 0.4)\n \n## What Didn't Work\n* Word based-label smoothing\n* Manifold mixup\n* Shift DA\n* Adding eye\n* Pseudo labels\n\n## Especially thanks to these kernels\n* [[LB 0.67] one pytorch transformer solution](https://www.kaggle.com/code/hengck23/lb-0-67-one-pytorch-transformer-solution) [@hengck23](https://www.kaggle.com/hengck23)\n* [GISLR [LB 0.63]: On the Shoulders](https://www.kaggle.com/code/roberthatch/gislr-lb-0-63-on-the-shoulders) [@roberthatch](https://www.kaggle.com/roberthatch)\n* [Animated Data Visualization](https://www.kaggle.com/code/danielpeshkov/animated-data-visualization) [@danielpeshkov](https://www.kaggle.com/danielpeshkov)", "2241984": "Hearty congratulations for the result and kudos to you for the approach note as well @donjyarahoi", "2242850": "donjyarahoi congratulations with great result!💥 Thanks for sharing - it helps to improve skills!🤝🙂" }, "source": "meta" }