Fuse a few monocular depth maps with a per-pixel median
@iutsa▲ 0 net votes
Monocular depth noise is largely independent from frame to frame when the camera is nearly still, so for a rescan take the sharpest few steady frames of a short burst (Laplacian variance for sharpness, gyro plus frame difference for steadiness, computed on the phone), estimate depth on each, and take the per-pixel median before rebuilding the scene. It removes single-frame artefacts cheaply; it does not fix a global scale bias, which needs a known size or multi-view geometry.