Working with intermediate pose representations for action recognition.
Notable papers:
[1] PoTion: Pose MoTion Representation for Action Recognition
[2] PA3D: Pose-Action 3D Machine for Video Recognition
[3] Simple yet efficient real-time pose-based action recognition
[4] Depth-Aware Action Recognition: Pose-Motion Encoding through Temporal Heatmaps
[5] Make Skeleton-based Action Recognition Model Smaller, Faster and Better
some older papers:
- https://arxiv.org/pdf/2104.13586v1.pdf
- https://arxiv.org/pdf/1907.13025.pdf
- https://arxiv.org/pdf/1703.03492.pdf
- https://arxiv.org/pdf/2107.12213v2.pdf
- https://arxiv.org/pdf/2108.07181v2.pdf
- https://arxiv.org/pdf/2108.04536v1.pdf
wacv 2022:
- https://openaccess.thecvf.com/content/WACV2022/papers/Shah_Pose_and_Joint-Aware_Action_Recognition_WACV_2022_paper.pdf
- https://openaccess.thecvf.com/content/WACV2022/html/Li_NUTA_Non-Uniform_Temporal_Aggregation_for_Action_Recognition_WACV_2022_paper.html
- https://openaccess.thecvf.com/content/WACV2022/html/Kataoka_Spatiotemporal_Initialization_for_3D_CNNs_With_Generated_Motion_Patterns_WACV_2022_paper.html
- https://openaccess.thecvf.com/content/WACV2022/html/Li_PERF-Net_Pose_Empowered_RGB-Flow_Net_WACV_2022_paper.html
TODO:
- 2048 filters may work, but for now takes too long to train