This is an official implementation of Lightweight Human Pose Estimation Using Loss Weighted by Target Heatmap that is accepted to ICPR TCAP workshop and honored as Best Paper Award (Honorable Mention). The code is developed based on the repository of MMPose.
In this work, we lighten the computation cost and parameters of the deconvolution head network in SimpleBaseline and introduce an attention mechanism that utilizes original, inter-level, and intra-level information to intensify the accuracy. Additionally, we propose a novel loss function called heatmap weighting loss, which generates weights for each pixel on the heatmap that makes the model more focused on keypoints. Experiments demonstrate our method achieves a balance between performance, resource volume, and inference speed.
| Input Size | #Params | GFLOPs | AP | AP50 | AP75 | APM | APL | AR |
|---|---|---|---|---|---|---|---|---|
| 256 × 192 | 3.1M | 0.58 | 65.8 | 87.7 | 74.1 | 62.6 | 72.4 | 72.1 |
| 384 × 288 | 3.1M | 1.30 | 69.9 | 88.8 | 77.5 | 66.0 | 76.7 | 75.5 |
Inference speed on Intel Core i7-10750H CPU and NVIDIA GTX 1650Ti (Notebooks) GPU.
| Model | AP | FPS(GPU) | FPS(CPU) | GFLOPs |
|---|---|---|---|---|
| MobileNetV2 | 64.6 | 57.6 | 19.3 | 1.59 |
| ShuffleNetV2 | 59.9 | 51.8 | 20.2 | 1.37 |
| ViPNAS | 67.8 | 22.6 | 4.6 | 0.69 |
| Lite-HRNet18 | 64.8 | 12.8 | 10.3 | 0.20 |
| Lite-HRNet30 | 67.2 | 7.5 | 6.2 | 0.31 |
| Ours | 65.8 | 55.2 | 18.3 | 0.58 |
Please refer to MMPose's documentation and add the file to corresponding direction.
The code is developed using python 3.7 on Ubuntu 20.04. The code is developed and tested using a single NVIDIA RTX 3090 GPU. Other platforms or GPU cards are not fully tested.
@inproceedings{li2022lightweight,
title={Lightweight Human Pose Estimation Using Loss Weighted by Target Heatmap},
author={Li, Shiqi and Xiang, Xiang},
booktitle={International Conference on Pattern Recognition},
pages={64--78},
year={2022},
organization={Springer}
}
Thanks to: