Awesome

Diffusion-Based 3D Human Pose Estimation with Multi-Hypothesis Aggregation [ICCV2023]

The PyTorch implementation for "Diffusion-Based 3D Human Pose Estimation with Multi-Hypothesis Aggregation" .

Dependencies

Make sure you have the following dependencies installed (python):

• pytorch >= 0.4.0
• matplotlib=3.1.0
• einops
• timm
• tensorboard

You should download MATLAB if you want to evaluate our model on MPI-INF-3DHP dataset.

Datasets

Our model is evaluated on Human3.6M and MPI-INF-3DHP datasets.

Human3.6M

We set up the Human3.6M dataset in the same way as VideoPose3D. You can download the processed data from here. data_2d_h36m_gt.npz is the ground truth of 2D keypoints. data_2d_h36m_cpn_ft_h36m_dbb.npz is the 2D keypoints obatined by CPN. data_3d_h36m.npz is the ground truth of 3D human joints. Put them in the ./data directory.

MPI-INF-3DHP

We set up the MPI-INF-3DHP dataset following P-STMO. However, our training/testing data is different from theirs. They train and evaluate on 3D poses scaled to the height of the universal skeleton used by Human3.6M (officially called "univ_annot3"), while we use the ground truth 3D poses (officially called "annot3"). The former does not guarantee that the reprojection (used by the proposed JPMA) of the rescaled 3D poses is consistent with the 2D inputs, while the latter does. You can download our processed data from here. Put them in the ./data directory.

Evaluating our models

You can download our pre-trained models, which are evaluated on Human3.6M (from here) and MPI-INF-3DHP (from here). Put them in the ./checkpoint directory.

Human3.6M

To evaluate our D3DP with JPMA using the 2D keypoints obtained by CPN as inputs, please run:

python main.py -k cpn_ft_h36m_dbb -c checkpoint -gpu 0 --nolog --evaluate h36m_best_epoch.bin -num_proposals 5 -sampling_timesteps 5 -b 4

You can balance efficiency and accuracy by adjusting -num_proposals (number of hypotheses) and -sampling_timesteps (number of iterations).

For visualization, please run:

python main_draw.py -k cpn_ft_h36m_dbb -b 2 -c checkpoint -gpu 0 --nolog --evaluate h36m_best_epoch.bin -num_proposals 5 -sampling_timesteps 5 --render --viz-subject S11 --viz-action SittingDown --viz-camera 1

The results will be saved in ./plot/h36m.

MPI-INF-3DHP

To evaluate our D3DP with JPMA using the ground truth 2D poses as inputs, please run:

python main_3dhp.py -c checkpoint -gpu 0 --nolog --evaluate 3dhp_best_epoch.bin -num_proposals 5 -sampling_timesteps 5 -b 4

After that, the predicted 3D poses under P-Best, P-Agg, J-Best, J-Agg settings are saved as four files (.mat) in ./checkpoint. To get the MPJPE, AUC, PCK metrics, you can evaluate the predictions by running a Matlab script ./3dhp_test/test_util/mpii_test_predictions_ori_py.m (you can change 'aggregation_mode' in line 29 to get results under different settings). Then, the evaluation results are saved in ./3dhp_test/test_util/mpii_3dhp_evaluation_sequencewise_ori_{setting name}_t{iteration index}.csv. You can manually average the three metrics in these files over six sequences to get the final results. An example is shown in ./3dhp_test/test_util/H20_K10/mpii_3dhp_evaluation_sequencewise_ori_J_Best_t10.csv.

Training from scratch

Human3.6M

To train our model using the 2D keypoints obtained by CPN as inputs, please run:

python main.py -k cpn_ft_h36m_dbb -c checkpoint/model_h36m -gpu 0 --nolog

MPI-INF-3DHP

To train our model using the ground truth 2D poses as inputs, please run:

python main_3dhp.py -c checkpoint/model_3dhp -gpu 0 --nolog

Testing on in-the-wild videos

We follow P-STMO to test our model on custom videos. You can use an off-the-shelf 2D keypoint detector (such as AlphaPose) to yield 2D poses from images and use our model to yield 3D poses. The 2D keypoint detectors are trained on COCO dataset, which defines the order of human joints in a different way from Human3.6M. Thus, our model needs to be re-trained to be compatible with the existing 2D detectors.

To evaluate our pre-trained model on in-the-wild videos, you can download in_the_wild_best_epoch.bin from here.

To train our model from scratch, you should download 2D keypoints of Human3.6M in COCO format from here. Our model take these 2D keypoints as inputs and outputs 3D joint positions in Human3.6M format. Then, you can use the following command to train our model:

python main_in_the_wild.py -k detectron_pt_coco -c checkpoint/model_in_the_wild -gpu 0 --nolog

After that, you can evaluate our models on in-the-wild videos using video-to-pose3D. Please follow the below instructions.

1. Follow their README.md to set up the code.
2. Put the checkpoint in_the_wild_best_epoch.bin to the ./checkpoint folder of their repo.
3. Put ./in_the_wild/videopose_diffusion.py to the root path of their repo.
4. Put other files in ./in_the_wild folder to the ./common folder of their repo.
5. Run videopose_diffusion.py!

Citation

If you find this repo useful, please consider citing our paper:

@InProceedings{Shan_2023_ICCV,
    author    = {Shan, Wenkang and Liu, Zhenhua and Zhang, Xinfeng and Wang, Zhao and Han, Kai and Wang, Shanshe and Ma, Siwei and Gao, Wen},
    title     = {Diffusion-Based 3D Human Pose Estimation with Multi-Hypothesis Aggregation},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {14761-14771}
}

Acknowledgement

Our code refers to the following repositories.

• VideoPose3D
• DiffusionDet
• MixSTE
• video-to-pose3D

We thank the authors for releasing their codes.