In this repository, we provide training and testing codes for 3DPW-Crowd (Table 5) and 3DPW (Table 8).
We use the pre-trained ResNet-50 weights of xiao2018simple to achieve faster convergence, but you can get the same result by training longer.
Download the file of weights and place it under ${ROOT}/tool/
.
Use the appropriate config file to reproduce results. For example, to reproduce 3DPW-Crowd (Table 5), run
python train.py --amp --continue --gpu 0-3 --cfg ../assets/yaml/3dpw_crowd.yml
Remove --continue
if you don't want to the use pre-trained ResNet-50 weights.
Add --exp_dir
argument to resume training.
Note: CUDA version may matter on the training time. Normally it takes 2hours per epoch when I used cuda-10.1. But when I use cuda-10.2, it takes 4~6hours per epoch. Pytorch version is 1.6.0.
Download the experiment directories from here and place them under ${ROOT}/output/
.
To evaluate on 3DPW-Crowd (Table 5), run
python test.py --gpu 0-3 --cfg ../assets/yaml/3dpw_crowd.yml --exp_dir ../output/exp_03-28_18:26 --test_epoch 6
To evaluate on 3DPW (Table 8), run
python test.py --gpu 0-3 --cfg ../assets/yaml/3dpw.yml --exp_dir ../output/exp_04-06_23:43 --test_epoch 10
You can replace the --exp_dir
with your own experiments.