[rfonod]

Summary

Two fixes to the feature detectors, both of which made a configured value mean something other than what it says.

• LoFTR input is padded to a multiple of 8. kornia's LoFTR matches on a coarse grid of stride 8 and maps grid cells back to pixels with the ratio of the input height to the coarse height. When a side of the processed frame is not a multiple of 8 that ratio is not 8, so every match carried an error proportional to its distance from the origin. _LoFTRMatcher.match now zero-pads both images at the bottom and right (new dl.pad_to_multiple, which leaves pixel coordinates unchanged) and drops matches that fall in the padding.
• max_features caps BRISK, KAZE and AKAZE. These detectors take a response threshold, not a count, and stabilo converted max_features into one with linear models fitted at native resolution (stabilo/thresholds/models/). The models ignored downsample_ratio, and AKAZE's went negative above about 11k keypoints, so the reference frame kept every keypoint. The detectors now run at their own threshold (brisk_threshold, kaze_threshold, akaze_threshold) and keep the max_features strongest by response (ref_multiplier x max_features in the reference frame), as ORB and SIFT already do. The thresholds default to OpenCV's (30, 0.001, 0.001); they were previously fallbacks that the shipped models made unreachable.

Output impact

• LoFTR: unchanged when both sides of the processed frame are multiples of 8 (no padding is applied). Otherwise the bias is removed. On a 4K aerial frame warped by a known homography, the mean corner error drops from 22.3 to 0.33 px at downsample_ratio 0.18 (691x389) and from 7.9 to 0.16 px at 0.25 (960x540).
• BRISK, KAZE, AKAZE: keypoint counts and homographies change. Reference/current keypoints on a 4K pair:

Setting	Detector	v1.4.4	v1.4.5
0.18 / 2000	BRISK	225 / 137	4000 / 2000
0.18 / 2000	AKAZE	70 / 40	1149 / 1171
1.0 / 8000	BRISK	26,706 / 6289	16,000 / 8000
1.0 / 8000	AKAZE	228,234 / 7471	16,000 / 8000

• ORB, SIFT, RootSIFT and the sparse learned detectors: unchanged. The default-path regression test passes.

Tests

• 12 cases for the cap: at most max_features / ref_multiplier x max_features at two downsample ratios, the kept keypoints and descriptors equal the uncapped detector's strongest, and a cap above the detected count keeps every keypoint.
• 4 offline cases for the padding (a stand-in LoFTR records the padded shapes and returns a match in the padding), and 2 end-to-end LoFTR cases recovering a known homography at sizes not divisible by 8.
• With the fixes reverted, 16 of the 18 new cases fail.
• Full suite: 205 passed, 18 skipped (CUDA and DL end-to-end); with STABILO_DL_TESTS=1, all 40 DL tests pass. stabilo CLI smoke test passes; the wheel and sdist build and pass twine check.

Docs

• docs/usage.md section 9 describes the cap for every classical detector and the LoFTR padding; the parameter table lists the new threshold defaults.
• stabilo/cfg/default.yaml, the README feature list and scripts/README.md no longer describe the threshold models as the source of the budget. scripts/find_threshold_models.py stays as a tool for choosing a threshold.

Other

• stabilo/thresholds/models/ is removed from the package; package-data now ships **/*.yaml only.
• Version bumped to 1.4.5 in __init__.py, CITATION.cff, .zenodo.json, and the README citation.