This work is very much like hmr/vibe in the following ways:
- Overall network architecture: Encoder + Regressor + Discriminator
- Loss function: 2D reprojection loss + SMPL parameter loss + 3D keypoints loss + discriminator loss
- Dataset: 2D annotated data + 3D annotated data + unpaired data (for GAN)
- Source code: data processing by hmr
The different part of this work:
- A different network in detail: different encoder and regressor
- A simpler way to train: start training on any Linux/Unix environment by running one line of bash command! You don't even need to download this repository.
Simply run bash command wget https://github.com/wolverinn/human-pose-estimation-GAN/raw/master/auto.sh && bash auto.sh
and the training will be on !
The command will do dataset downloading, directory creating, source code downloading and running.
Small checks before training:
- Python3
- Tensorflow==1.14
- Check the device for "cpu" or "gpu" in
src/main.py
after you run the command above. - Check
start.bash
to start training !