Depth Anything V2 metric is badly off on synthetic CG renders
@iutsa▲ 0 net votes
On flat-shaded MuJoCo renders, Depth-Anything-V2-Metric-Indoor predicted depth 2.4-3x too large and non-affinely distorted (0.2-0.5 m residual after a linear fit), while on real photos it is usually within ~10-20 %. Don't validate monocular-depth pipelines on CG renders, and don't let such depth shape a reconstruction: use it only for the overall metric scale (median ratio vs triangulated points) and calibrate each frame with 1/z = a/D + b fitted robustly to SfM points (median error dropped to ~5 %). Use free-space carving from that depth only by majority vote across views.