TASK 22 — Same-label duplicate suppression in ObjectMap

A third of the scene graph was redundant: one physical object represented by several nodes. This task measures the causes, ships the fix for the tractable half, and records why the other half was deferred.

Why the existing dedup could not fire

ObjectMap had two mechanisms, both keyed on 3D IoU of axis-aligned boxes:

• merge in add (same label): IoU >= 0.3 or centre <= 0.4 m
• cross-label NMS in finalize: IoU >= 0.5

Over 14 scenes, node pairs whose centres are within 0.5 m have median IoU 0.055, and none of 691 reach NMS_IOU — the cross-label rule has never fired in this dataset. MERGE_IOU contributes on 12% of pairs; the 0.4 m distance clause carries the rest.

This is not a lifter accuracy problem. Those pairs are close: their box surfaces sit a median 3 cm apart (53% under 5 cm, 82% under 15 cm), and the boxes are thin slabs — median extents 0.55 × 0.32 × 0.09 m, aspect 7.6:1. A LiDAR sweep sees one face of an object, so two viewpoints produce adjacent, disjoint surface patches. Two boxes of ~0.01 m³ whose centres are 0.43 m apart cannot overlap, and IoU is a step function below overlap: 3 cm apart and 3 m apart both score exactly 0.

TASK 21 made this worse, predictably — in arabic_room the median IoU between co-located nodes fell 0.158 → 0.010 once boxes tightened. The thresholds were implicitly calibrated against bloated boxes.

What shipped

box_gap() — shortest distance between two AABB surfaces, 0.0 when touching. Unlike IoU it stays informative below overlap.

finalize suppresses the weaker of two co-located nodes on either the legacy IoU >= NMS_IOU (any label) or centre <= NMS_DIST and box_gap <= NMS_GAP, restricted to identical labels. Best-supported node wins: most observations, then score.

NMS_DIST = 0.4 m matches MERGE_DIST deliberately. add already merges same-label nodes inside that radius — but a node's centre moves as it accumulates points, so two nodes created further apart can drift inside it with nothing re-checking. This is that check, deferred until the centres settle, which also bounds what it can recover.

threshold	redundant	counting MAE	mAP@1	R@1	P@1
off	216	2.8356	0.2134	0.2552	0.3253
d0.4 g0.05	204	2.8126	0.2123	0.2524	0.3267
d0.4 g0.15	204	2.8126	0.2123	0.2524	0.3267
d0.3 g0.05	215	2.8351	0.2131	0.2546	0.3251

Counting MAE improves on 6 scenes, worsens on 1, unchanged on 7. Dropping to 0.3 m removes the effect entirely. The gap term is not binding at 0.4 m (0.05 and 0.15 are byte-identical) and is kept only as a guard against a large box whose centre coincides with a small one.

What was deferred

Cross-label duplicates are the larger half — 82% of duplicated GT objects — and a hard gate rather than a threshold: add skips any candidate whose label differs before evaluating distance or IoU. Restricting the new rule to identical labels was a deliberate call: applying it across labels would also collapse genuinely touching distinct objects (a pillow on its sofa) and would hide detector label instability behind whichever label won.

Full analysis, the synonym-vs-misclassification split, and two candidate approaches are recorded in docs/tasks/backlog.md B3.

Changes

file	change
src/xiao_hei_vln/perception/object_map.py	box_gap(); _suppresses(); NMS_DIST / NMS_GAP; absorbed_labels in to_list()
perception_benchmark/replay_score.py	--nms-dist, --nms-gap
perception_benchmark/box_quality.py	dup_objects / redundant columns
tests/test_object_map.py	5 tests for the suppression rule
docs/tasks/backlog.md	B3

absorbed_labels records a suppressed label on the survivor rather than discarding it. It is dormant while the rule is same-label-only, and is the hook B3 needs so label instability stays visible instead of being swallowed.

One implementation trap worth noting: export() builds a throwaway view that shares the same _Node objects as the live map, so recording absorbed labels on the nodes would leak into the live map and accumulate every tick. They are held in a per-map dict keyed by node_id instead, and test_export_does_not_mutate_the_live_map pins that.