We propose a novel approach HANet that utilizes the keypoints’ kinematic features, following the laws of physics. Our method addresses temporal issues with these proposed features, effectively mitigates the jitter, and becomes robust to occlusion.
We propose a hierarchical transformer encoder that incorporates multi-scale spatio-temporal attention. We use multi-scale feature maps, i.e., leverage all layers’ attention maps, and improve performance on benchmarks that provide sparse supervision.
We propose online mutual learning that enables joint optimization between refined input poses and final poses, which chooses an online learning target by their training losses.
We conduct extensive experiments on large datasets and demonstrate that our framework improves performance on tasks: 2D pose estimation, 3D pose estimation, body mesh recovery, and sparsely-annotated multi-human 2D pose estimation.
Clone the repo:
https://github.com/KyungMinJin/HANet.git
Install the HANet requirements using conda
:
# conda
conda create env --name HANet python=3.6
conda activate HANet
pip install -r requirements.txt
Sub-JHMDB data used in our experiment can be downloaded here. Refer to Official DeciWatch Repository for more details about the data arrangement.
Dataset | Pose Estimator | 3D Pose | 2D Pose | SMPL |
---|---|---|---|---|
Sub-JHMDB | SimpleBaseline | ✔ |
Note that datasets should be downloaded and prepared before training.
Run the commands below to start training on Sub-JHMDB:
python train.py --cfg configs/config_jhmdb_simplebaseline_2D.yaml --dataset_name jhmdb --estimator simplebaseline --body_representation 2D
Results on 2D Pose:
Dataset | Estimator | PCK 0.05 (Input/Output):arrow_up: | PCK 0.1 (Input/Output):arrow_up: | PCK 0.2 (Input/Output):arrow_up: | Checkpoint |
---|---|---|---|---|---|
Sub-JHMDB | simplebaseline | 57.3%/91.9% | 81.6%/98.3% | 93.9%/99.6% | Google Drive |
Results on 3D Pose:
Dataset | Estimator | MPJPE (Input/Output):arrow_down: | Accel (Input/Output):arrow_down: |
---|---|---|---|
Human3.6M | FCN | 54.6/52.8 | 19.2/1.4 |
Human3.6M | Mhformer | 38.3/35.4 | 0.8/0.8 |
3DPW | PARE | 78.9/77.1 | 6.9/6.8 |
AIST++ | SPIN | 107.7/69.2 | 5.7/5.4 |
We prepare all visualization codes as soon as possible.
Visualize comparison on Sub-JHMDB
Visualize comparison on AIST++
Visualize comparison on 3DPW
Visualize comparison on AIST++
@inproceedings{jin2023kinematic,
title={Kinematic-aware Hierarchical Attention Network for Human Pose Estimation in Videos},
author={Jin, Kyung-Min and Lim, Byoung-Sung and Lee, Gun-Hee and Kang, Tae-Kyung and Lee, Seong-Whan},
booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
pages={5725--5734},
year={2023}
}