We provide two learning-based post-processing methods deciwatch and smoothnet for speed-up and smoothing respectively.
We provide the config files for DeciWatch: DeciWatch: A Simple Baseline for 10x Efficient 2D and 3D Pose Estimation.
@article{zeng2022deciwatch,
title={DeciWatch: A Simple Baseline for 10x Efficient 2D and 3D Pose Estimation},
author={Zeng, Ailing and Ju, Xuan and Yang, Lei and Gao, Ruiyuan and Zhu, Xizhou and Dai, Bo and Xu, Qiang},
journal={arXiv preprint arXiv:2203.08713},
year={2022}
}We use checkpoints trained on SPIN-3DPW for demo speed up. Checkpoints with different intervals and q values are provided. If you need more checkpoints trained on various datasets and backbones, please refer to the official implementation of DeciWatch.
| Interval | Window Q | Config | Download | Speed Up | Precision Improvement (MPJPE In/Out) |
|---|---|---|---|---|---|
| 10 | 1 | deciwatch_interval10_q1 | model | 10X | 99.35 / 95.85 |
| 10 | 2 | deciwatch_interval10_q2 | model | 10X | 99.45 / 96.37 |
| 10 | 3 | deciwatch_interval10_q3 | model | 10X | 99.60 / 96.98 |
| 10 | 4 | deciwatch_interval10_q4 | model | 10X | 99.58 / 96.87 |
| 10 | 5 | deciwatch_interval10_q5 | model | 10X | 99.78 / 97.39 |
| 5 | 1 | deciwatch_interval5_q1 | model | 5X | 99.31 / 95.05 |
| 5 | 2 | deciwatch_interval5_q2 | model | 5X | 99.35 / 95.05 |
| 5 | 3 | deciwatch_interval5_q3 | model | 5X | 99.45 / 94.84 |
| 5 | 4 | deciwatch_interval5_q4 | model | 5X | 99.45 / 94.94 |
| 5 | 5 | deciwatch_interval5_q5 | model | 5X | 99.55 / 94.48 |
To use different settings of DeciWatch in demo, specify --speed_up_type with the checkpoint name. For example, you may use --speed_up_type deciwatch_interval10_q3 for 10X speed up with a window size of 31. Simply set --speed_up_type deciwatch to use default setting deciwatch_interval5_q3. The meaning of interval and q can be found in the original paper.
We provide the config files for SmoothNet: A Plug-and-Play Network for Refining Human Poses in Videos.
@article{zeng2021smoothnet,
title={SmoothNet: A Plug-and-Play Network for Refining Human Poses in Videos},
author={Zeng, Ailing and Yang, Lei and Ju, Xuan and Li, Jiefeng and Wang, Jianyi and Xu, Qiang},
journal={arXiv preprint arXiv:2112.13715},
year={2021}
}We use checkpoints trained on SPIN-3DPW for demo pose smoothing. Checkpoints with different window size are provided.
| Window Size | Config | Download | Precision Improvement (MPJPE In/Out) | Smoothness Improvement (Accel In/Out) |
|---|---|---|---|---|
| 8 | smoothnet_windowsize8 | model | 96.85 / 95.84 | 34.62 / 7.13 |
| 16 | smoothnet_windowsize16 | model | 96.85 / 95.61 | 34.62 / 6.35 |
| 32 | smoothnet_windowsize32 | model | 96.85 / 95.03 | 34.62 / 6.11 |
| 64 | smoothnet_windowsize64 | model | 96.85 / 95.26 | 34.62 / 6.02 |
To use different settings of SmoothNet in demo, specify --smooth_type with the checkpoint name. For example, you may use --smooth_type smoothnet_windowsize8 for pose smoothing with a window size of 8. Simply set --mooth_type smoothnet to use default setting smoothnet_windowsize8. The meaning of windowsize can be found in the original paper.