TASK 36 — Viewpoint redundancy, the dwell bias, and a flat-object extent estimator

Why

After the merge gate collapsed each carpet to one node ([[TASK 32]]), the right carpet's box (node #13) was seen to shrink and drift as observations accumulated: GT-coverage peaked at 53% (14 obs) and decayed to 20% (240 obs), with the box migrating into the −y/+x corner.

Root cause, measured:

• Not _core_points — it discarded 0.0% of points and clustering fired on 0/240 observations (the carpet cloud is clean and coplanar).
• The estimator + viewpoint redundancy. Each dimension of a node's box is the point-count-weighted mean of every view's own max−min extent. Of node #13's 240 observations, 228 (88% of the weight) came from a single 0.5 m robot dwell cell, each seeing only a thin 0.42 m slice. A handful of earlier pass-through views saw the full 1.35 m width (≈ GT 1.44) but at ~3% weight were drowned out. So the box averages down to the dwell view and its centroid.

The estimator has no defense against viewpoint redundancy — it weights by points/observations collected, not by how much new information a viewpoint adds. Standing still degrades the box.

The design (two complementary levers)

1. Extent estimator — prefer reach over average (implemented, this task). For flat nodes (median per-view z-extent ≤ FLAT_Z_M), size the box from the max per-view extent in x/y — the reach of the single best view that saw the whole object — instead of the weighted mean. z stays the weighted mean (~0 regardless). Volumetric nodes are untouched: their mean cancels the ~6× inflation of pooling near-face LiDAR slabs, which max would instead amplify.

2. Capture-time novelty gate (implemented, responder._inject_visible). The tick is decoupled so the ScanAccumulator ingests LiDAR every tick (cloud stays dense/registered) while perception (detect → lift → ObjectMap.add) fires only on a novel viewpoint. With a 360° panorama the frame content depends only on position, so novelty is pure translation:

kept = KDTree of accepted (x, y) positions   # ALL previous, not just the last
novel(pos) := distance-to-nearest-kept > NOVEL_DIST   (~0.3–0.5 m)

Comparing against all kept positions (not just the last) is required so a revisit/loop doesn't re-admit an already-covered pose. This stops dwell over-counting at the source and cuts perception compute. It is object-agnostic, so it caps redundant frames but does not by itself grow the box — hence it pairs with lever 1.

Prototype results (offline, arabic_room, GT-coverage %)

	right #32	left #58
baseline (weighted mean, no gate)	20	64
flat-max estimator, no gate (shipped)	56	87
gate 0.3 m + mean	56	64
gate 0.3 m + p90	67	76
gate 0.3 m + max	69	77

Notes: p90 alone (no gate) fails on the right carpet (15%) because 95% of its views are the narrow dwell slice — only max escapes extreme redundancy. Once the gate rebalances the view distribution, the safer high-percentile works. Gating slightly hurts the already-diverse left carpet (87→77) by discarding useful views — evidence the two levers should be tuned together, not stacked blindly.

Shipped here

Estimator — object_map._Node._recompute, flat-conditional (median z ≤ FLAT_Z_M), two levers: - Max extent (xy): the reach of the single best view, not the dwell mean. Carpet GT-coverage 20→56% (right), 64→87% (left). - Union-midpoint centre (xy): the midpoint of the per-view boxes' union instead of the point-weighted centroid, so a box built from partial reach isn't dragged onto the seen half. Right carpet centre err 0.52→0.24 m (coverage 69→82%), left 0.30→0.12 m (77→84%).

Volumetric nodes are untouched (weighted mean, which cancels the ~6× near-face pooling inflation max/union would amplify). Why flat-only, not generic: only coplanar clouds have no near/far-face ambiguity, so their extreme edges are real surface points; a volumetric object's extremes are offset near-face slabs, so an extremal statistic (max/union-midpoint) trusts the most wrong view.

Novelty gate — responder._inject_visible: scan accumulator runs every tick; detect → lift → fuse is gated on _viewpoint_is_novel, a nearest-neighbour check over ALL kept (x,y) (revisits caught) against novel_viewpoint_m. Wired through _PerceptionSettings as XIAO_HEI_NOVEL_VIEWPOINT_M. Reproducible offline: replay_score.py/dump_debug.py --novel-viewpoint-m <m>. Default is 0.3 m (ON) — flipped from off after the box-quality gain and the ~11× compute saving were judged to outweigh the small recall cost on the live robot (the accumulator still ingests LiDAR every tick, so the cloud stays dense).

Measured (arabic_room, frozen dets @0.6, flat estimator on)

gate	perceived	pred	mAP@0.5	mAP@1.0	IoU@0.25
off (0.0)	397	29	0.221	0.263	0.051
0.3 m (shipped)	36	24	0.199	0.235	0.091

The estimator alone (gate off, vs the pre-TASK-36 mean/centroid baseline) is a clean win: mAP@0.5 0.195→0.221 (+14%), centre_err 0.240→0.221 m, no recall cost. The gate then trades ~11× compute + IoU@0.25 +80% for a small mAP drop.

Tests: test_flat_node_box_takes_max_extent_not_mean, test_flat_node_centre_is_union_midpoint_not_weighted_centroid, test_volumetric_node_still_uses_mean_extent, test_novelty_gate_skips_perception_until_robot_moves, test_novelty_gate_disabled_perceives_every_tick.

Open / next

• Sweep novel_viewpoint_m and max vs high-percentile across the corpus against the real metric; max is safe on coplanar clouds but fragile to a single spill view. Relates to [[B9]] (box containment) and B6.
• Recover the gate's box-quality gain without the recall cost — gate the fusion weight rather than the observation (perceive every frame, down-weight redundant viewpoints in _recompute). Keeps recall while killing the dwell bias.
• The score/SAM detection thresholds this pairs with are [[TASK 37]].