Improve detection accuracy in the warehouse world

[adharshvenkat]

Motivation

Detection quality in the synthetic warehouse is rough: borderline confidence scores, some prompts not matching the world's meshes well, and depth sampling that's weak for large objects.

Scope

• box_threshold=0.30 (semantic_detector.py) is low; borderline detections (e.g. observed door confidence ~0.34-0.52) suggest raising it or tuning per-label thresholds.
• Generic DEFAULT_PROMPTS vocabulary may not match the actual meshes in the warehouse world; validate/tune prompt wording against what GroundingDINO actually detects well in this scene.
• Depth sampling (projection.py::sample_depth) uses a small fixed window around the bbox center; for large objects (e.g. a shelf spanning much of the frame) this can still land on a background gap. Consider sampling across a fraction of the bbox interior instead of a fixed small window.
• No outlier rejection when multiple observations are fused for a single instance (related to Semantic map fuses multiple instances of the same label into one incorrect pose #6, may be implemented together).

Acceptance criteria

• Reduced rate of low-confidence/borderline detections in a standard warehouse pass.
• Depth sampling validated against at least one large object (shelf/pallet) known to previously misproject.
• Document final threshold/prompt choices and why, so they're not silently re-broken later.