How cos526-project-ss-yz/cos526-project works
Project type and purpose
• This is a research‐grade Python pipeline for “Zero‐Shot Point Cloud Completion with 3D Diffusion Priors.”
• It ingests partial 3D scans (Redwood, ShapeNet, nuScenes), preprocesses them, runs a zero‐shot completion using a pretrained diffusion model (Hunyuan3D-v2), and evaluates/visualizes results.
File structure
Top level:
• Scripts & notebooks:
– prepare_data.py, preparedata_nuscenes.py – data ingestion & preprocessing
– shape_completion_pipeline_nuscenes.py – core completion pipeline
– shape_completion_pipeline_demo_*.ipynb – interactive demos (Redwood, ShapeNet)
– register_results.ipynb, camera_vis.ipynb, densityplotcheck.py – evaluation & visualization
• Data folders:
– for_inputs/… – raw partial scans (masks & world planes)
– for_GT/… – ground‐truth masks & transformations
– frames_data_full/… – downloaded full Redwood frames (subfolder per sample)
• Utils:
– utils/depth_utils.py, chamfer_python.py, register_utils.py, gd_register_utils.py – common math, registration, geometry helpers
• Environment & configuration:
– create_env.sh, .gitignore, README.md
README for additional insights
• Prerequisites: PyTorch, diffusers, transformers, Open3D, PYG, plus Hunyuan3D-v2 as an external dependency
• Usage steps: download data, run prepare_data.py, then launch the notebooks to perform zero‐shot completion on different datasets.
Elements to include for clarity
– Blue for data storage
– Green for processing scripts
– Orange for model inference
– Purple for utilities
– Yellow for notebooks/visualization
Detail
• Every essential artifact (scripts, data folders, utils, external dependencies)
• Don’t over‐complicate—use the file tree to enumerate components
• Keep the flow linear and labeled, reflecting exactly how you’d run prepare_data.py → pipeline → notebooks