AI3D-388 · Fence detection channel — passes 1-3 on the A1 train split (2026-09-07)

2026-09-07 · technical peer · branch t3code/implement-fence-detection · train split only

New independent fence evidence channel in the guardrail detector (branch t3code/implement-fence-detection, ticket AI3D-388). Master switch fence_detection_enabled, default false.
With the switch off the outputs are byte-identical to the v0.5.0 baseline — guardrails.json, XML modulo CreationDate, all PNGs; verified on segments 002, 060, 086.
Data: A1 export with the scan-angle filter off and number_of_returns + RGB present (/mnt/d/a123-data/Abschnitt_1_no_angle_filter_returns). Split reused from the AI3D-387 greenery work (train 64 / val 64 / test 14, dev/analysis/fence/split.json). Everything here is train only; val/test untouched.
Pass 1 (measure). Channel with no post-gates on 11 train segments → 68 candidates, all labelled visually by two independent Opus judges plus a tie-break: 24 vehicles, 16 stolen guardrails, 13 trees, 10 hedges, 5 misc, 0 fences. Both judges also found no uncovered fence in those segments (woodland-lined motorway).
Pass 2 (gates from the labelled negatives). 6 physically motivated post-gates reject all 68. A scan of the remaining 53 train segments (263 candidates) left 13 survivors — among them the first real fences.
Pass 3. Height-resolved returns/colour (the fence sheet above the grass line is what gets judged), wall-adjacency and vehicle rules, and a crest-flatness cue for transmissive-but-not-green sheets (a wire mesh is itself 2-return). On the 12 re-detected segments: 90 candidates → 11 accepted = 6 fence-like true positives (5 distinct objects: a 34 m 1.8 m mesh fence, a mesh panel, a 2 m mesh fence in a gore, a wooden post-and-rail fence in 2 fragments) + 1 solid 2.4 m panel wall (taxonomy) + 5 structure false positives (bridge parapet, greenhouse/solar eave and roofs, a building facade).
Stolen guardrails: zero after the gates — all 16 pass-1 rail refits are rejected by guardrail_overlap/rail_like_top. The guardrail list itself is untouched by design: the channel runs after the rail gates and never through them.
Final sweep + val gate (new). Code 78d825c over all 64 train and all 64 val segments, 0 failures: train 331 candidates → 12 accepted, val 418 → 9 accepted. Of the 9 val survivors only 1-2 are fence-like and both are judge-disputed; 2 are rest-area vehicles and 5-6 are structures. Zero guardrails stolen on val. Precision is not deployable yet — see §5.
Open. Fences are rare in A1 (5 objects in 64 train segments); a full train+val sweep with the final code is running. Taxonomy questions for HBW (post-and-rail fence vs Geländer, solid panel wall vs Fence §7.6). Far-range recall limit: a low fence at 45-55 m outboard is ~15 pts/m and outside the 35 m corridor.

1. What was built

File	Role
`guardrails/_model_fence.py`	Config slice: every `fence_*` key, JSON parity with the rest of the model.
`guardrails/fence_evidence.py`	O(grid) accumulator: 0.25 m cells, 0.3-3.0 m band histogram in 0.1 m bins, low band 0.10-0.30 m, tall band 3-8 m, multi-return and green (ExG > 0.06) counts whole-band and upper-band (hag ≥ 1.0 m), raw 3D moments, returns/RGB availability flags.
`guardrails/fence.py`	Cell gates, fit through `fence_view_config` → shared `detect_instances`, support cells, candidate assembly.
`guardrails/fence_metrics.py`	23+ metrics: `top_p90`/`top_spread`, `tall_frac`, `low_band_frac`, `multi_return_rate(_upper)`, `green_rate(_upper)`, `interior_frac`, `edge_distance_med_m`, `nearest_guardrail/wall_lateral_m`, planarity/linearity.
`guardrails/fence_gates.py`	The 8 ordered post-gates (table below).
`guardrails.json + segment_<seg>_fence_candidates.json`	Output: a `"fences"` list plus a sidecar with every candidate and its `accepted`/`reject_reason` — fully auditable, replayable offline.
`colors.py / export.py`	Fences drawn red, display ids `Fnn`, legend row painted only when a fence is present; XML feature type `"fence"`.
`tests/test_fence_channel.py, test_fence_gates.py, test_fence_upper_band.py`	536 tests passing.
`dev/analysis/fence/`	`render_fence_candidates.py` (the 6-panel close-ups), `aggregate_candidates.py`, `replay_gates.py` (offline what-if on the candidate JSONs), `labels_*.json`, `split.json`.

2. Post-gates — order, rule, rationale

Applied in order inside detect_fence_candidates; the first firing rule is the recorded reject_reason. Defaults are shown inline in the rule column.

#	Gate	Rule (default)	Why that number
1	`inside_carriageway`	`interior_frac > 0.50`	All 24 labelled vehicles read `interior_frac` 1.0; hedges/trees read 0.0 and rails never exceed 0.5. A fence is never on the carriageway.
2	`guardrail_overlap`	nearest rail `< 1.00 m`	The 16 guardrail-labelled candidates are the segment’s own accepted rail refitted by the fence grid, 0.0-0.8 m from it.
3	`wall_overlap`	nearest wall `< 3.00 m`	The wall channel already owns bridge parapets, overpass edges and noise walls; pass-2 survivors sat 0.15 m and 2.78 m from an accepted wall.
4	`rail_like_top`	`top_p90 < 1.20 m`	Labelled rails and bare post rows measure `top_p90` 0.95-1.15 m; the fence prior starts at 1.0-2.5 m.
5	`tall_mass`	`tall_frac > 0.10`	Trees carry 0.11-0.71 of their mass above 3 m and the two noise-wall sheets 0.11/0.14; every other candidate is ≤ 0.08.
6	`vehicle_like`	`length < 20 m` and edge dist `< 7 m` and `multi < 0.05`	A short opaque body on a road surface the lane XML does not model. Costs a fence class and says so: short fences inside the clear zone are sacrificed.
7	`vegetation`	(`green > 0.40` and `multi > 0.12`) or (`multi > 0.55` and `top_spread > 0.50 m`)	Upper-band rates preferred, whole-band fallback, `null` keeps. A wire mesh is itself 2-return, so transmissivity alone would delete real fences (092/3 reads multi_up 0.67) — the crest separates the classes: fence-like objects measure `top_spread` 0.02-0.30 m, the brown shrub row 059/4 measures 1.32 m.
8	`sparse`	`points_per_m < 40`	Loose guard against dust and stray-return bridges only; the sparsest labelled real object is 67 pts/m. Never a class discriminator.

Pass-1 replay: all 68 labelled negatives rejected

replay_gates.py over the pass-1 candidate JSONs with the current defaults. No fence existed in those 11 segments, so a perfect score here is a necessary, not sufficient, condition.

Label	n	Rejected by
vehicle	24	23 `inside_carriageway` + 1 `vegetation`
guardrail	16	15 `guardrail_overlap` + 1 `rail_like_top`
tree	13	13 `tall_mass`
hedge	10	2 `guardrail_overlap`, 2 `rail_like_top`, 6 `vegetation`
noise_wall	2	1 `inside_carriageway` + 1 `tall_mass`
other	2	2 `inside_carriageway`
building	1	1 `inside_carriageway`
fence	0	— none existed in the 11 pass-1 segments

3. Pass-3 accepted candidates (11 of 90 on 12 segments)

Metrics from segment_<seg>_fence_candidates.json (accepted=true). multi_up/green_up are the upper-band (hag ≥ 1.0 m) rates. 048/4 and 048/6 are two fragments of one physical fence, so 6 fence-like rows = 5 distinct objects.

seg/id	verdict	Fable label	Opus label	len m	top_p90 m	multi_up	green_up	spread m	edge m
092/3	fence	fence (0.90)	fence (0.85)	34.4	1.75	0.67	0.15	0.10	21.4
048/4	fence	fence (0.85)	fence (0.90)	15.8	2.30	0.08	0.25	1.25	6.7
048/6	fence	fence (0.85)	fence (0.90)	9.2	1.33	0.13	0.37	0.19	6.8
008/6	fence	fence (0.70)	—	19.7	1.99	0.91	0.39	0.30	17.1
092/8	fence	fence (0.75)	flagged in scan as probable fence	9.5	1.75	0.93	0.40	0.02	22.7
127/1	taxonomy	noise_wall (0.70)	noise_wall (0.55)	37.7	2.37	0.02	0.15	0.29	18.3
077/2	disputed	unclear (0.50) — wire fence	building_or_structure (0.90) — greenhouse eave	12.5	1.30	0.50	0.01	0.19	30.9
039/5	false pos	building_or_structure (0.85)	building_or_structure (0.88)	8.7	2.94	0.05	0.03	0.08	18.0
077/3	false pos	building_or_structure (0.90)	building_or_structure (0.92)	9.5	2.46	0.20	0.34	1.30	24.8
077/4	false pos	building_or_structure (0.90)	building_or_structure (0.92)	9.5	2.48	0.18	0.37	0.87	24.0
008/11	false pos	building_or_structure (0.80)	building_or_structure (0.85)	11.4	2.82	0.08	0.06	0.24	18.6

077/2 is a genuine dispute, recorded as such. Fable read a 1.3 m wire fence with posts every ~2.5 m; the Opus judge read the eave of a greenhouse/solar-panel roof slab (an inclined plane attached at the line in the cross-section). It is counted as a false positive in the tally above, i.e. the pessimistic reading. Its green_up of 0.01 with multi_up 0.50 is exactly the signature the crest cue was built to keep, which is why it survives either way.

4. Figures

Click any image for the full-resolution 1800×900 render; ←/→ (or A/D, swipe) cycle within one card. Each 6-panel render is: oblique RGB · oblique by number_of_returns · cross-section · top-down RGB · top-down returns · along-candidate side view.

True positives: fences found on train

Six panels each: oblique RGB, oblique coloured by number_of_returns (grey 1, orange 2, red 3+), cross-section, top-down RGB, top-down returns, along-candidate side view.

092/3 · 1.8 m mesh fence, 34 m · dead-straight top wire, posts every ~3 m, grass at the base · multi_up 0.67, green_up 0.15, top_spread 0.10 m, top_p90 1.75 m

048/4 · wooden post-and-rail fence · three rails at 0.5/0.85/1.15 m on posts every ~2 m, single-return; the adjacent sign board lifts top_p90 to 2.30 m and top_spread to 1.25 m

008/6 · 2 m mesh fence in the gore between carriageway and ramp · thin flat-topped sheet, ~0.3 m wide in the cross-section · multi_up 0.91, green_up 0.39, top_spread 0.30 m

092/8 · mesh panel 1.75 m under the tree canopy · perfectly rectangular side view with two posts · multi_up 0.93, green_up 0.40, top_spread 0.02 m, tall_frac 0.0002 (no canopy mass)

Taxonomy question + disputed

Both survive every gate; what they should be called is the open part.

127/1 · solid 2.4 m painted panel wall with posters, 38 m · uniformly single-return (multi_up 0.02), filled rectangle in the side view, no post gaps · below the wall channel’s 2.5 m floor: fence or low noise wall?

077/2 · disputed · Fable read a 1.3 m wire fence with posts every ~2.5 m; Opus read the eave of a greenhouse/solar roof slab · multi_up 0.50, green_up 0.01, top_spread 0.19 m, edge dist 30.9 m · counted as a false positive in the tally

Residual false positives (structures)

What the current gate set still lets through: opaque man-made sheets that are not fences.

039/5 · overbridge parapet · opaque grey planar slab, tall_frac 0.097 just under the 0.10 ceiling, accepted noise wall 4.5 m away (just over the 3.0 m wall_overlap clearance)

077/3 · gabled roof profiles of the same farm structure · the side view is a sawtooth of inclined planes rising to 2.5 m, not posts with a flat top · top_spread 1.30 m, multi_up 0.20

008/11 · facade/eave of a small building with a tiled roof · top-down is a solid 12x8 m rectangle, cross-section a multi-layer mass · top_p90 2.82 m, multi_up 0.08, green_up 0.06

What the gates remove (pass-1 negatives)

One example per dominant reject reason, from the 68 labelled pass-1 candidates and the scan.

002/3 · car · rejected by inside_carriageway: interior_frac 1.00 · solid slab in the cross-section, zero multi-return

002/0 · hedge · rejected by vegetation (AND branch): green 0.63 with multi-return 0.39 · blobby cross-section, ragged crest

018/0 · stolen guardrail — the segment’s own accepted rail refitted by the fence grid over 83 m · rejected by guardrail_overlap: 0.0 m from the accepted rail, top_p90 0.95 m (also below the 1.20 m rail_like_top floor)

059/4 · brown non-green shrub row 30 m outboard · transmissive but not green, so the colour branch cannot see it · rejected by the crest cue: multi_up 0.80 with top_spread 1.32 m (fences measure 0.02-0.30 m)

Whole-segment views

Context for 092 (the clearest true positive) and the far-range recall limit.

092 overview · all 0.3-3 m evidence cells coloured by number_of_returns; accepted candidate polylines red, rejected grey · the accepted F03 runs along the property boundary at 34.6 m offside

092 top-down product overlay · fences red (F03, F08), w_beam guardrails green, supports purple · the fence list is additive: the guardrail products are byte-identical to the channel-off run

008 probe · recall limit · a low fence at 45-55 m outboard, ~15 pts/m, outside the 35 m lateral corridor · never proposed as a candidate, so no gate can be blamed for it

Val gate survivors (final sweep)

The 9 candidates the final gate set accepts on the 64 val segments, plus the one new train survivor. Val was never used for tuning.

val 015/0 · <b>disputed</b> · Fable: 1.2 m post-and-rail fence, ladder of regular posts with a top and a lower rail; Opus: double-rail steel rest-area barrier (the upper rail ramps into an end terminal) · len 28.6 m, top_p90 1.25 m, multi_up 0.05, top_spread 0.20 m

val 049/4 · <b>disputed</b> · Fable: gabled building roof; Opus: mesh panel fence on an embankment (flat top stepping down in panel-length increments) · len 8.8 m, top_p90 1.95 m, multi_up 0.44, green_up 0.29

val 093/4 · structure FP · rows of tilted solar panels on posts — parallel inclined slabs at 4-5 m carried on columns · top_p90 2.60 m, top_spread 0.72 m: mass at height beside the fitted line, which no current metric measures

val 120/2 · structure FP · Fable: shelter/pergola at a parking bay; Opus: a vehicle — not a fence either way · razor-flat top at 2.45 m, single-return, green_up 0.64

val 015/6 · vehicle FP · parked car at the rest area, unmistakable silhouette with wheel voids · edge distance 8.3 m, just above the 7 m vehicle_like bound, so the rule cannot see it

train 129/1 · <b>new true positive</b> · 1.3 m mesh fence, 27 m, posts every ~4-5 m, mesh 2-return between them · multi_up 0.63, green_up 0.24, top_spread 0.28 m · both judges agree

5. Final-code sweep over all train + val segments (val gate)

Code 78d825c (all gates), run from a frozen worktree over every train (64) and every val (64) segment, 0 failures. The val segments were never used for tuning — no threshold in §2 has ever seen them.

Reject reason	train	val
`inside_carriageway`	145	202
`guardrail_overlap`	80	85
`wall_overlap`	4	1
`rail_like_top`	8	12
`tall_mass`	44	71
`vehicle_like`	1	4
`vegetation`	37	34
rejected	319	409
accepted	12	9
candidates	331	418

Train: 12 survivors

The 11 of §3 plus segment 129 candidate 1 — a 1.3 m mesh fence 27 m long with posts every ~4-5 m, called a fence by both judges (Fable 0.85, Opus 0.94). See card q6.

Val: 9 survivors

Metrics from final/val/branch_000/<seg>/segment_<seg>/segment_<seg>_fence_candidates.json. Labels from labels_final_fable.json and labels_final_opus.json; where the two judges disagree, both readings are shown.

seg/id	verdict	Fable label	Opus label	len m	top_p90 m	multi_up	green_up	spread m	edge m
015/0	disputed	fence (0.90) — post-and-rail	guardrail (0.60) — double-rail barrier	28.6	1.25	0.05	0.13	0.20	16.0
049/4	disputed	building_or_structure (0.85) — roof	fence (0.78) — mesh panel	8.8	1.95	0.44	0.29	0.31	23.4
015/5	vehicle	vehicle (0.90) — parked HGV	vehicle (0.75) — HGV trailer flank	12.7	2.95	0.01	0.62	0.70	7.9
015/6	vehicle	vehicle (0.95) — car	vehicle (0.93) — car	11.6	1.65	0.00	0.05	0.60	8.3
093/4	structure	building_or_structure (0.90) — solar rows	building_or_structure (0.78) — panels on columns	18.6	2.60	0.08	0.40	0.72	25.0
094/5	structure	building_or_structure (0.90) — solar rows	building_or_structure (0.80) — roofed deck	9.0	2.39	0.11	0.33	1.28	28.5
120/1	structure	building_or_structure (0.80) — underpass abutment	vehicle (0.85)	29.0	2.85	0.01	0.53	0.95	6.8
120/2	structure	building_or_structure (0.85) — shelter/pergola	vehicle (0.75)	9.6	2.90	0.01	0.64	0.57	20.1
120/3	structure	building_or_structure (0.85) — same shelter	vehicle (0.75)	9.2	2.84	0.01	0.63	0.49	20.5

Val verdict, plainly

Of the 9 accepted, 1-2 are fence-like and both of those are disputed (015/0, 049/4) — the two judges swap sides on them: Fable calls 015/0 a fence and 049/4 a roof, Opus calls 015/0 a guardrail and 049/4 a fence.
2 are rest-area vehicles (015/5 HGV, 015/6 car) at edge distance 7.9 m and 8.3 m — just above the 7 m vehicle_like bound, so the rule cannot reach them.
5-6 are structures: solar-panel rows (093/4, 094/5), a shelter/pergola in two fragments (120/2, 120/3), an underpass abutment (120/1). Fable and Opus disagree on 120/1-3 (structure vs vehicle) — not a fence either way.
Zero guardrails stolen on val. 85 rail refits rejected by guardrail_overlap and 12 by rail_like_top; the guardrail list is untouched by construction, since the channel runs after the rail gates and never through them.
Precision on val is therefore not deployable yet. The residual family is one thing: man-made structures with mass at height beside the fitted line — roof, canopy and panel planes. No current metric measures that. See §7.1.

Val recall notes Opus judge only

Reported by the Opus judge from the overviews and RGB overlays; not independently verified by Fable. Recorded as leads, not as counted misses.

093 candidate 0 — 103 m at offset −26.8 m, bounded top 1.70 m, rejected as vegetation (multi 0.92, green 0.51). Geometrically fence-like; may be an overgrown fence inside the hedgerow.
094 — a ~50-60 m thin mostly-2-return line ~30 m outboard was never proposed as a candidate.
120 — a fence along the solar-farm boundary, 30-45 m outboard, running the whole segment, never proposed: beyond or at the 35 m corridor edge.
Coverage is routinely truncated: the fences in 129, 049 and 048 all visibly continue past the fitted fragments.

6. Method notes

Measurement first. No threshold was written before the 68 labelled pass-1 negatives existed; every default in the gate table cites the measured spread it came from.
Two independent vision judges + tie-break per segment (Opus judges in pass 1, Opus scan + Fable in pass 3). Disagreements are recorded, not smoothed — see 077/2.
Gates are replayable offline on the candidate JSONs: uv run python dev/analysis/fence/replay_gates.py --runs-root /home/ai/gr_fence_out/p1 --labels dev/analysis/fence/labels_p1.json [--set fence_min_top_p90_m=1.4].
The fence channel is deliberately not routed through the precision gate G2 or the edge gate E1 — both ask “is this near a carriageway edge”, which a boundary fence legitimately is not.
Off-state byte-identity check is part of the acceptance, not an afterthought.
Cost: peak RSS 3-6 GB per segment, O(grid) evidence only — never point lists.

7. Next steps

Pass-4 structure cue. An O(grid) lateral-occupancy-at-height metric: mass 1-3 m to the side of the fitted line, above 1 m hag, single-return. Roofs, canopies and panel rows all have it; a fence sheet has nothing beside it. This is the single change that addresses the whole residual false-positive family.
Vehicle rule. Raise the edge-distance bound, or drive it from a lane-XML parking / rest-area zone — the two val vehicle FPs sit at 7.9 m and 8.3 m, just outside the current 7 m bound.
Post-cadence metric from the side-view evidence: periodic ground-to-top posts, which roofs and parapets do not have.
Coverage. A fence-view merge/bridging step so one continuous physical fence becomes one instance (048, 049 and 129 are all truncated or split today).
HBW taxonomy. Post-and-rail fence vs Geländer; solid panel wall vs Fence §7.6; double-rail barriers (the 015/0 dispute is exactly this question).
Corridor. Fences at 30-45 m outboard need fence_offset_max_m > 35 plus a sparse-fit path; today they are never proposed, so no gate can recover them.