Observed arrival · 2026-09-01
Label Refinery’s self-labeling construction-site loop
A technically explicit perception project that uses geometry, learned models, and synthetic construction scenes to generate and improve its own training labels.
Field notes
The documented loop preserves each pipeline result as an immutable dataset version, allowing planning to read what the previous round recorded rather than overwrite it. Its first stage uses LiDAR sweeps and joint states to discover unnamed instances; later stages add class names and retrain a detector. The evaluation setup uses a synthetic 60-second construction scene with a held-out oracle, while Grounding DINO supplies names only after geometry has proposed the objects.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue