Ambiguity-Aware studies the ambiguity problem in the task of unsupervised 3D human pose estimation from 2D counterpart, please refer to ICCV2022 for more details.
conda create -n uvhpe python=3.6
conda activate uvhpe
pip install -r requirements.txt
# for output, tensorboard, visualization
mkdir log output vis models data
Download our preprocessed dataset into data and pretrained models into models from webpage
This part will be updated soon.
We put some samples with preprocessed 2d keypoints at scripts/demo_input. Run inference with command sh demo.sh and output can be found at scripts/demo_output.
- baseline
python main.py --cfg ../cfg/h36m_gt_adv.yaml --pretrain ../models/adv.pth.tar --gpu 0 --eval - scale
python main.py --cfg ../cfg/h36m_gt_scale.yaml --pretrain ../models/tmc_klbone.pth.tar --eval --gpu 0
- baseline
python main.py --cfg ../cfg/pre_adv.yaml --pretrain ../models/pre_adv.pth.tar --gpu 0 --eval - scale
python main.py --cfg ../cfg/pre_tmc_klbone.yaml --pretrain ../models/pre_tmc_klbone.pth.tar --gpu 0 --eval
Note: baseline is our reproduced version fo "Unsupervised 3d pose estimation with geometric self-supervision"
use the pretrained model from Human3.6M
python eval_lsp.py --cfg ../cfg/h36m_gt_scale.yaml --pretrain ../models/tmc_klbone.pth.tar
The expected MPJPE and P-MPJPE results on Human36M dataset are shown here:
| Input | Model | MPJPE | PMPJPE |
|---|---|---|---|
| GT | baseline | 105.0 | 46.0 |
| GT | best | 87.85 | 42.0 |
| Pre | baseline | 113.3 | 54.9 |
| Pre | best | 93.1 | 52.3 |
Note: MPJPE from the evaluation is slightly different from the performance we release in the paper. This is because MPJPE in the paper is the best MPJPE during training process.
-
Using ground-truth 2D as inputs:
baseline
python main.py --cfg ../cfg/h36m_gt_adv.yaml --gpu 0best
python main.py --cfg ../cfg/h36m_gt_scale.yaml --gpu 0 -
Using predicted 2D as inputs:
baseline
python main.py --cfg ../cfg/pre_adv.yaml --gpu 0best
python main.py --cfg ../cfg/pre_tmc_klbone.yaml --gpu 0
