Handoff: Tree Recall on the No-Angle-Filter Dataset

2026-09-05 · AI3D-373 · for the next agent session · generated by Claude (Fable 5)

Summary

Continue tree-recall work in the verticalsigns detector. The previous round (v6 hedge gate) is shipped (spot recall 0.920, 2/26 FP on truth_v4). This round exploits a new, denser dataset generated without the scan-angle filter — trees are much more visible (1.62× points in seg 021) and a new number_of_returns field arrives per point. Goal unchanged: improve tree recall until perfect, until more data is needed, or until human input is required. Miro answers no more truth questions — iterate autonomously.

The new dataset

/mnt/d/a123-data/Abschnitt_1_no_angle_filter_returns/ (184 GB, treat as read-only source — stage what you need, budget disk accordingly).

No scan-angle filter: segment 021 has 1.62× the points of the old staged data (42.6M vs 26.3M). Previously-missing trunks/crowns may now be present.
New per-point key: number_of_returns (uint8) alongside the old keys (points float64, scan_angle int8, intensity/red/green/blue uint16). Multi-return points are a strong vegetation signal — foliage multiplies returns, hard surfaces don't.
Layout: branch_000/lane_points/segment_NNN/*_run3_points.npz (131 segments) and branch_001/… (11 segments). Note the extra branch_NNN/ level vs the old staged root /home/ai/seg3d_stage_260807b/A1B0/lane_points/segment_NNN/.
run3_geoshift.json is byte-identical to A1B0's (x 694541.44…, y 5386372.37…, z 467.41…) — truth-spot coordinates transfer directly, no reprojection needed.

Current state (already done)

Worktree /home/ai/dev/wt-vs-recall, branch feat/AI3D-373-tree-recall, PR #7 → master (Bitbucket), HEAD ad41427, all pushed, 1597 tests green.
v6 gate: size-based hedge disjunction — legacy (area ≥ 250 m², relief ≤ 1.5) OR huge (area ≥ 1250 m², relief ≤ 2.2) OR ext (area ≥ 450 m² ∧ len ≥ 44 m ∧ relief ≤ 1.68), all under shared conjuncts + h_top ≤ 3.3. Implemented as BigBandBranches in tree_instances.py; knobs tile_hedge_gate_* + hedge_band_as_medium_vegetation. Repo defaults all None; live values in /home/ai/veg373_work/recall/phase3/cfg/v6.json and v6mv.json.
Eval on truth_v4 (113 spots, 87 tree / 26 not): v6 recall 0.920, FP 2/26. Baselines: v0 0.460/1FP, v4 0.920/5FP, v5 0.897/1FP.
Corpus (104 prod-paired segments): prod 14.0M tree pts → v6 23.9M (1.71×, zero losses vs prod).
Demo videos delivered on N: (260716_videos/260831_tree_recall_v6/).
Full context in memory files tree-recall-diagnosis.md, miro-hedge-definition.md, ai3d-373-vegetation-split.md under /home/ai/.claude/projects/-home-ai-dev-3dai-iolabs-pointcloud-verticalsigns/memory/.

Where the remaining recall lives (ranked)

Rerun on the new data, re-score. cheapest first
The known-unfixable FNs died on RF point mass (~90-pt trunks: 066 c15 tiles s10/s15, 083 c3 s1). 1.6× density may fix them with zero code change. Run this experiment first.
number_of_returns as a feature. Candidate RF feature or pre-filter (multi-return fraction per cluster/tile). Pursue only if lead 1 leaves FNs on the table; changing the RF feature set means retraining/threshold work — check with Miro first.
096 c6 sub-parent hedge strip — below parent granularity; likely still needs splitter work; reassess on denser data.

How to run

Detector: from wt-vs-recall, uv run the detect entry point with a cfg from phase3/cfg/ (copy phase3/run_v6.sh as a template). The detector expects the A1B0-style layout (lane_points/segment_NNN/), so stage per-branch or symlink branch_000/lane_points.
Eval: phase3/eval/ scripts score tree_clusters.csv against phase3/phase0/truth_v4.json (spot-based; coordinates valid on new data — same geoshift). Expected-gate checker: phase3/gate3/check_v6_expected.py.
Truth caveat on new data: point counts per spot change; recall by spot-hit stays valid, weighted-by-points comparisons across datasets do not.

Rules violations caused real damage before

Soft labels: 021 c0, 068 c1, 121 c5, 108 c34, 066 c8, 094 c1, 066 c15 — Miro: "it can be either… tree trunks are often missing". Either verdict counts as correct; ambiguity default = tree. Tune thresholds only against hard-labeled objects.
One --out dir per detector process. Corridor post-passes glob and rewrite every segment_*/verticalsigns.json in --out; concurrent sharing is a data race (bit us twice: out/v6_corpus masks untrusted, out/v5_corpus non-tree CSVs rewritten). ≤4 concurrent processes, each with its own --out.
RAM ≤ 32 GB/process: stream per-file, float32 after applying the float64 geoshift.
Runtime budget < 20 min/segment (051 and 015 already exceed it — pre-existing).
Scratch under /home/ai/veg373_work/ (/tmp is tmpfs, wiped on WSL reboot).
Keep candidates.min_height_m at its current value; CSV_FIELDS is append-only.
Scope: vegetation + vertical signs in this repo only (no guardrails, no seg3d colour knobs, no sign path).
Worktrees /home/ai/dev/3dai.iolabs.pointcloud.verticalsigns (main checkout), wt-seg3d-fuse-prod-260827, and others belong to other sessions — work only in wt-vs-recall (or a fresh worktree). Bare git stash/pop forbidden (shared stack); use git stash push -u -m "<tag>" + apply by SHA if unavoidable.
Commits: 1-line, AI3D-373: …, no co-author lines. Push to the existing branch/PR. Ruff only via uv tool run ruff.

Open items you inherit

verify out/v6_corpus/*/verticalsigns.json possibly race-contaminated — verify or regenerate before any seg3d-fuse consumes them (tree_clusters.csv + eval numbers unaffected).
Proposed A1 production config (v6 gate + medveg emission) delivered to Miro; adoption is his call.
PR #7 description stale (no Bitbucket API creds on this machine; pushes still update commits).
Future notes, not to iterate now: sub-shared gate max_relief inversion footgun (documented, low), 032 c0 stems 1.609 knife-edge (unlabeled).

Definition of done for this round

A rerun + rescore on the new dataset with a v6-vs-new comparison table (spot-based recall + FP count), and each remaining FN either fixed, reclassified as soft, or documented with the specific evidence for why it needs data/annotation the detector doesn't have. Report deltas to Miro extremely concisely.

Markdown twin of this handoff: /home/ai/veg373_work/recall/HANDOFF_no_angle_filter.md