ShootFlow's cull command groups the frames of a burst so you review one keeper instead of five near-copies. Earlier today it learned to match zooms and wide pans. That left an odd gap: on held-out photos, pans of 6% grouped only 85-86% of the time, while 8% pans grouped 96-100%.
The note that found the gap guessed at a cause. The cull hashes 25 windows of 88% of each frame on a 5 x 5 grid of offsets, 3% apart. A 6% pan moves the content 6.4% of the frame, and the guess was that this falls "between the grid's steps". The proposed fix was a 7 x 7 grid: 24 more hashes per frame, and every catalogue analysed again.
Ask the misses first
Before building the finer grid, this run took every 6% pan pair the cull missed and sorted it by the reason. The sorter is tools/cull_crop_gate_tune.py --why, run on pairs cut from the 331 real JPEGs of a public test-image repository. Of 993 pairs, 901 grouped, and all 92 misses failed the same way:
- the windows lined up fine: for 89 of the 92, one frame's centre window was within 0-4 bits of a window of the other (out of 64);
- but the whole frames were 25-33 bits apart, and the window comparison only ran for pairs within 24.
So the grid did its job. 6.4% is about one pixel from the 6% offset at the analysis size. A pre-check in front of it (the "gate") never let these pairs reach the windows. The wide-pan pass added this morning is gated at 44, but it only looks at the larger offsets, so nothing compared these pairs at all.
The fix, and its price
The crop pass's gate is now 44, the same as the wide-pan pass's. Nothing else changed: the window threshold and the required gain stay as they were. Picked on the tune half of the photos and scored once on the held-out half:
| before | after | |
|---|---|---|
| held-out burst-pair recall | 0.985 | 0.991 |
| 6% pans grouped (horizontal, diagonal) | 0.86, 0.85 | 1.00, 1.00 |
| false merges in 13,695 pairs of different photos | 26 | 26 |
The price is speed. With the gate at 44, the window check runs for nearly every pair in the all-pairs scan, and done the old way it doubled the grouping time on 2,000 frames (8.6 s to 17.8 s). So the crop and wide-pan checks are now one packed comparison. Each frame's 49 crop window pairs and 56 wide-pan pairs sit side by side in one 6,720-bit integer. One XOR compares them all, and a lane-wise popcount checks each lane against its own limit. The last step of the popcount is a single multiply, which sums each lane's byte counts into its top byte. The result: 10.7 s instead of 8.6 s, about 25% more, with unit tests that pin the packed check to the plain distances on 1,500 random pairs.
What is still open
10% pans still group 86-95% of the time. These misses are real window misses. Their best window pair is 6-11 bits apart, just past the limit of 5, because an 11.1% content shift falls 0.9% short of the grid's 12% offset. That is where a finer grid would earn its re-analysis. The 6% case was not. The full numbers, the tuning grid and one deviation from the usual pick rule are in eval/cull-crop-gate-2026-11-18.md.
The pairs are cut from single exposures, so subject motion and parallax between real burst frames are not in this test. A real shoot's picks remain the check that matters.