Two changes, big payoff: AAB0 #0 went from 243/250 to 250/250 byte-exact on the captured Wine enrollment frame. Average across 9 AAB0s improved from ~94% to ~98%. 1. FILL was 128 (= the Q0 mid-gray byte value). DLL pads its 3×3 tiles with Q16 mid-gray = 0x800000 (= 8388608). Using 128 propagated through the doh convolution chain to non-padding pixels and changed NMS results in EDGE TILES (rows 2 and cols 2 of the 3×3 grid). With the fix, my port's tile 8 is byte-exact with captured F250 tile 8. 2. Corner-tile tightening in _a960_passes_global_edge: for tile (i=GRID-1, j=GRID-1) — bottom-right — cap gy at oy + last_row_step (= 102 exclusive for 112-px frame). Without this, my port keeps 7 tile-8 kps with gy in [102, 108] that DLL drops. Empirically derived from 9-AAB0 enrollment capture; the kp orientation of these kps and the per-AAB0 gy_max in DLL's tile 8 (range 82-101) suggests the actual rule may be more nuanced (per-frame variable), but the exclusive `gy < oy + step` empirically catches most edge cases for the captured data. Per-AAB0 byte-exact (gx, gy) match after fix: #0: 250/250 #1: 239/250 #2: 239/250 #3: 243/250 #4: 244/250 uunicorn#5: 248/250 uunicorn#6: 249/250 uunicorn#7: 246/250 uunicorn#8: 246/250 Residual drift (1-11 per frame) is concentrated in TILE 8 (corner) where DLL applies a per-frame gy cut whose rule isn't yet pinned down. For example AAB0 #1 DLL keeps only 3 of ~12 high-resp tile-8 kps; DLL's kept kps have gy ≤ 82 while my port's only_mine kps have gy in [89, 101]. The cut isn't strict top-N or fixed gy threshold. Diagnostic tools added: dev/find_edge_bound.py - per-tile bound aggregator across all AAB0s (lists observed gx/gy ranges) dev/try_per_tile_bound.py - tests a per-tile bound filter against my port's output Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

No license specified #5

Description

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions