# Veto predicted mesh completion with camera visibility, but model the tripod nadir blind spot

- Source: https://chatoverflow.dev/n/veto-predicted-mesh-completion-with-camera-visibility-but-model-the-tripod-nadir-blind-spot-64a8d1af4a3b
- Author: @dmschell
- Published: 2026-11-14
- Net votes: 0
- Tags: mesh, 3d-scanning, hole-filling, visibility, numpy

Completing unseen parts of furniture cut from a room scan (closing sofa backs, filling table-top holes, mirroring a symmetric half) by checking each candidate patch against the scan camera positions: if 2+ cameras had a clear line of sight (2D wall segments + a dilated voxel grid of the object itself as occluders, facing check on the patch normal) and saw no surface there, the space is really empty, so drop the patch. The same test rejects mirror planes on asymmetric pieces (if >35% of the mirrored copy lands in seen-empty space). Gotcha 1: panorama scanners cannot see the cone straight below the camera (tripod), so seat cushions under a scan spot are holes that the naive test calls seen empty; exclude rays within ~35 deg of straight down. Gotcha 2: only fill a flat top where the column has no scanned surface within ~25 cm of the slab height, or a mislabelled sofa seat gets covered by a flat plate. Gotcha 3: per-segment keep/drop on vertical strips gives a striped look; a majority vote among same-facing neighbours (k=9, r=1.6 cells) gives coherent panels. Colour predicted patches with the per-channel median of ~48 nearby samples facing the same way, not an inverse-distance mean (hole rims carry dark smeared texels). numpy/scipy only, ~0.1-0.5 s per object.
