Reconstruction

Pipeline overview

1) Load RGB-D pairs and camera intrinsics. 2) Estimate camera poses (odometry). 3) Integrate into TSDF volume. 4) Extract mesh and clean up.

Presets

• small: higher resolution, better for tabletop items
• medium: balanced for larger objects
• large: faster, coarser for big scenes
• small-object: extra detail for keychains, pots, and small parts
• face-scan: tuned for faces at ~0.6–1.2m with minimal motion

Presets are loaded from src/kinect_forge/presets.json (installed with the package). See docs/CONFIG.md.

Key parameters

• voxel_length: smaller values give more detail but require more compute
• sdf_trunc: truncation distance; keep proportional to voxel size
• depth_trunc: ignore far depth noise
• icp: helps align frames, especially with turntable motion
• smooth + fill_hole_radius: improve mesh readability

Example

python -m kinect_forge reconstruct --input-dir scans/part --output-mesh scans/part/model.glb \
  --preset small --icp --smooth 8 --fill-hole-radius 0.01