Python / PyTorch / RGB-D perception

PoseLoop

Instance detection to 6D pose estimation through hash-bound masks, manifests and symmetry-aware development evaluation.

Public source · Merged into main

Work and recorded results

Integrated Mask R-CNN and FoundationPose for RGB-D/CAD evaluation, completing 820 registrations on 25 development frames; detector F1 0.726 and joint F1 0.606 use a custom protocol.
Implemented deterministic mask-to-pose manifests, per-instance diagnostics and memory-bounded feature scoring that preserves full-candidate attention.

Historical outputs were reconstructed from frozen evidence; this repair did not rerun the GPU models. Registration completion is not pose accuracy.
The already-consumed development protocol is not official BOP evaluation, unseen-scene generalization or production robotics acceptance. The published hash transcription error has an explicit erratum.

Evidence source revision: e0f4e88a2395451cd978c6cf526e66511ae87c95

Updated 2026-09-14. CV wording and this page use the same project manifest.