How shubham-goel/4d-humans works
Project type and purpose
• This is a research‐oriented, full-pipeline 3D human reconstruction and tracking system (4DHumans/HMR2.0).
• It provides:
• A demo/inference API (images & videos)
• A training pipeline (data download, Hydra‐configured experiments, PyTorch‐Lightning)
• An evaluation suite for multiple benchmarks
• A simple Gradio front-end
File structure
• Top‐level scripts:
• demo.py (image mode inference)
• track.py (video tracking via PHALP)
• gradio_app.py (UI)
• train.py (training entry point)
• eval.py (evaluation entry point)
• Core python package: hmr2/
• configs & configs_hydra (Hydra‐based experiment/config management)
• datasets (data loaders, preprocessors)
• models (backbones, transformer components, heads, losses)
• utils (download, geometry, rendering, logging)
• Deployment/build: environment.yml, setup.py
README for additional insights
• Uses: PyTorch, PyTorch-Lightning, Hydra, Detectron2 (ViTDet), PHALP, SMPL model
• Workflow stages: data fetch, preprocessing, training, inference/demo, evaluation
• Outputs: .obj meshes, side/front renders, tracked 3D pose files