put your .obj files under ./data/image/resin/{lion}/, and run .py under ./code/: superv_01_data_01_mesh.py -> superv_01_data_02_focus.py -> superv_01_data_03_split.py
OR: direct unzip data.zip
check code in superv.py, comment or uncomment 'infer/train' lines. (the infer code is fully compatible to official TripoSR, the tarin code is configured for small network.)
logic correct, no distributed version, no large scale train.