- Create a new conda env from environment.yml:
conda env create -f environment.yml - Activate the environment:
conda activate escape_env
For training and evaluating the networks in ESCAPE, we require the backbone network predictions on HP3D, MPII, PW3D.
Using our preprocessed data:
- Download data ready to use with ESCAPE: GDrive
- Extract the data and modify config.cnet_dataset_path in config.py to point to the folder
- Download the trained CNet and RCNet checkpoints: GDrive
- Extract and place at ./ckpts
Modify options in config.py to select experiments, change backbones, and modify other parameters.
Run experiments.py to execute the tasks in config.tasks: python3 experiments.py
Important Options:
- config.tasks: is a list which controls what experiments will be performed
- config.backbone: changes the pre-trained backbone estimator being used
- config.test_adapt: enables intensive test-time adaptation, else only fast correction will be applied
- config.TTT_e_thresh: enables Energy-based sample selection, instead of adapting to all samples
- config.energy_thresh: changes the Energy threshold used for Energy-based sample selection