BAE · RGBD pose refinement

Reproduce the rgbd benchmark: DA3 → joint photometric D → fixed-structure E. Inspect the raw proposal and the held-out guard result on identical RGBD inputs.

Pinned protocol: DA3-GIANT-1.1 · 504 px · scene-wide sampling up to 192 views · 16 neighbors · 512 samples · 5,000 structure points/view · self-calibration · seed 0. Source

ScanNet++ bundle ZIP: colmap.zip + resized_images.zip

Or load an existing benchmark prediction.npz

Upload only data you are licensed to use. Inputs and results are temporary runtime files; they are never committed to this repository. COLMAP poses are used only for evaluation. The Space starts empty; no licensed dataset is distributed with it.

Prediction cache

Refinement

Benchmark metrics and guard verdict

Inspect

Guarded result Raw proposal Optimization history

History progress (%)

0 100

RGBD alignment before / after

Both clouds use the same depth, sampled pixels and self-calibrated intrinsics. Only the poses change. Display sampling does not affect optimization. AUC values are fractions; higher is better. Guard acceptance measures photometric consistency, not ground-truth pose improvement.

Selected input views

Download prediction, raw/guarded poses, history, metrics and provenance

Prediction provenance