Code repository for our solution to the EgoBody Challenge held by HBHA workshop at ECCV 2022 .
In this repository, we explore some effective data augmentations to improve the model generalization of 3D human pose and shape estimation from a single egocentric image.
Our code is mainly based on SPIN, please refer to this repository for the installation of the environment.
You need to download the following data to start the experiment:
Then you need to specify their paths in config.py
.
You also need to generate the 2D keypoints for calculating 2D joint loss by running:
python keypoints.py
The generated 2D keypoints data will save as .npy file for easy loading.
You can train on the EgoBody dataset using pre-trained model by running:
python train.py --name exp_name --pretrained_checkpoint=/path/to/pre-tained/model.pt
The checkpoints and tensorboard files will be saved in the logs
directory by default.
Please refer to the train_options.py
for adding more data augmentations and setting other parameters.
You can download our best model here.
The majority of this repository is borrowed from SPIN. We also use some functions from EFT and EgoBody. Thank these authors for their great work.