3D pose prediction with multiple cameras and bundle adjustment
To create the necessary conda environment, run
conda env create -f requirements.yml
The above will create a conda environment called 3d-pose, which you should conda activate before proceeding.
We appreciate any kind of feedback. Open an issue or send a message, and we'll try our best to implement that feedback
The script reconstruct_points.py is where the reconstruction happens. It takes as input a path to a config file, e.g. ./configs/ibl.yaml. To perform reconstruction run the code as follows:
python reconstruct_points.py --config ./configs/ibl.yaml --output_dir ./outputs/ --downsample 10 --reproj_thresh 6
The output_dir argument specifies where the outputs are saved. downsample specifices the rate to downsample the frames before saving to video. This does NOT downsample the 3D points, only the frames for visualization. reproj_thresh specifies a threshold for selecting frames with high reprojection error.
For a new dataset, make a new config file that follows the pattern of the ones in ./configs. This requires specifying paths to the videos in each view, CSV files in each view, camera parameters (intrinsics and extrinsics), and bodypart information (bodypart names, colors, etc.).