Video2Mesh

Video2Mesh turns a scan video into layered 3D assets: 3DGS visual scene, semantic object masks, object meshes, collision proxies, physics metadata and simulator adapters.

Read the published docs at relumeow.top/video2mesh/. The source Markdown remains in docs/.

Quick Start

cd /root/autodl-tmp/workspace/Video2Mesh
source /etc/network_turbo >/dev/null 2>&1 || true

bash run.sh dataset/<video>.mp4

bash run.sh wraps tools/run_video2mesh_quick.sh; if no argument is passed it uses VIDEO=/path/to/video.mp4 or the first video found under dataset/ or inputs/.

Current default route:

video
  -> COLMAP poses and full point cloud
  -> GraphDECO 3DGS
  -> GroundingDINO bbox prompts, with SAM/OpenCV auto prompts as fallback
  -> SAM2 masks
  -> 2D-to-3D semantic fusion
  -> COLMAP dense Delaunay scene collider mesh
  -> mesh semantic transfer and semantic object mesh splitting
  -> object mesh / completion jobs
  -> collider and physics proxies
  -> MuJoCo / Unity / Isaac assets

For commands, QA and research decisions, start here:

• Project Overview
• Pipeline And Commands
• Research Roadmap
• Mesh, Interaction And Completion
• Operations And Showcase
• Site And Remote Control

Generated data, exports, checkpoints, videos and model weights are intentionally ignored by Git.