mirror of
https://github.com/opencv/opencv.git
synced 2026-07-21 19:33:03 +04:00
Matt Van Horn
5e6a591357
imgproc: fix INTER_NEAREST_EXACT to match Pillow (5.x retarget of #28661) #28873 ## Summary Fixes `INTER_NEAREST_EXACT` producing different results from Pillow for images with dimensions that are multiples of 64 (e.g., 128, 192). Retargets #28661 to 5.x per core team request. ## Root cause The old 16-bit fixed-point arithmetic (`ifx`, `ifx0`, `>> 16`) rounds differently from Pillow's pixel-center coordinate mapping. At exact pixel boundaries (e.g., y=7 with 128 to 160: `0.4 + 7*0.8 = 5.999999999999999`), Pillow's truncation drops to 5 while the old opencv formula rounds to 6. ## Changes - `resize.cpp`: Replace `ifx/ifx0/ify/ify0` 16-bit fixed-point with an int64 arithmetic formula equivalent to `floor((i + 0.5) * src / dst)`, computed as `((int64_t)(i * 2 + 1) * src) / (dst * 2)`. This matches Pillow's pixel-center mapping exactly, is deterministic across platforms, and avoids FP rounding divergence entirely. - `test_resize_bitexact.cpp`: Add `NearestExact_PillowCompat` covering 4 dimension pairs including the reproducer from #28429. ## Testing - All 3 `Resize_Bitexact` tests pass (Linear8U, Nearest8U, NearestExact_PillowCompat) on this branch built against 5.x - Verified bit-exact match to PIL output (Python 3.14, Pillow) on 49 dimension combinations including random pairs - Built on macOS ARM64 with HAL (carotene/KleidiCV) active Fixes #28429