A 3D reconstruction lab whose point is keeping the viewer honest

1 minute read

Published:

A 3D reconstruction lab whose point is keeping the viewer honest

Lidar3D is an early research lab for feed-forward streaming 3D reconstruction: an ordered RGB or LiDAR stream becomes a camera trajectory, dense metric depth and a fused colored point cloud. Before anything else, the honest framing: it replays baked artifacts, it is not real-time and not SLAM by default, and the outdoor model is vendored, not mine. The heavy engine runs offline on a GPU and the public site is a static SPA that replays the result.

What is actually wired is a from-scratch depth-and-pose network (ResNet-18 backbone, a Siamese SE(3) pose head, aleatoric depth) with a real held-out trajectory error, reconstructing eight real indoor scenes from TUM, 7-Scenes and ICL. Alongside it, a genuinely vendored 2026 model handles four outdoor scenes and is kept labelled as vendored, plus Open3D ICP LiDAR odometry and a CPU synthetic engine.

The differentiator is not the network. It is a strict renderer-honesty discipline:

Own depth+pose net, held-out on TUM:
trajectory error (ATE) ~0.28 m · depth AbsRel 0.38 to 0.22 · ~12.8 M params
Four renderers (three.js, deck.gl, surfels, Potree LOD) draw the same baked cloud under one unified (x, minus-y, minus-z) transform. Where a renderer cannot do per-frame replay, the app approximates and says so in the UI; a metric without ground truth reads "none" with a reason.

The reason I built it this way is that reconstruction demos are usually the least honest part of the pipeline: a polished cloud that hides where the pose drifted, which numbers had no ground truth, or that two viewers are quietly showing different data. Here a deck.gl coordinate-mirror bug was found and fixed precisely because the four renderers are held to one transform, so a discrepancy is visible instead of pretty.

It does not claim to beat the state of the art, there is no textured mesh or Gaussian-splat output, and loop closure and global bundle adjustment are opt-in. The value is the discipline and the negative-results ledger, not a fidelity claim. Live · source.