{"full_name":"obss/sahi","owner":"obss","name":"sahi","shard":"1","url":"https://github.com/obss/sahi","description":"Framework agnostic sliced/tiled inference + interactive ui + error analysis plots","homepage":"https://obss.github.io/sahi/","topics":["object-detection","instance-segmentation","computer-vision","small-object-detection","large-image","pytorch","python","coco","deep-learning","machine-learning","remote-sensing","huggingface","fiftyone","satellite","tiling"],"tag_pages":["computer-vision","deep-learning","huggingface","instance-segmentation","machine-learning","object-detection","python","pytorch"],"language":"Python","license":"MIT","avatar":"https://avatars.githubusercontent.com/u/63360722?v=4","created_at":"2021-01-30T12:54:53Z","pushed_at":"2026-11-01T01:43:33Z","metrics":{"stars":5604,"forks":793,"watchers":47,"issues":0,"prs":891,"releases":120,"issues_open":0,"issues_closed":0,"prs_open":14,"prs_closed":86,"prs_merged":791,"discussions":392},"growth":{"day":{"stars":1,"forks":0,"base_day":"2026-11-15"}},"history":{"dates":["2026-11-10","2026-11-11","2026-11-12","2026-11-13","2026-11-14","2026-11-15","2026-11-16"],"stars":[5596,5599,5600,5601,5603,5603,5604],"forks":[793,793,793,793,793,793,793],"watchers":[null,47,47,47,47,47,47],"issues_open":[null,0,0,0,0,0,0],"prs_open":[null,2,2,14,14,14,14]},"boards":{},"same_owner":[],"issues":[],"discussions":[{"number":1508,"title":"YOLOVAL: a two-image sliced-inference check with game-specific ONNX models (paid)","comments":0,"category":"Showcase","created_at":"2026-11-04T18:44:54Z"},{"number":1476,"title":"Responsible disclosure: requesting a private channel to report a security issue","comments":2,"category":"General","created_at":"2026-10-15T04:38:56Z"},{"number":1420,"title":"🚀 SAHI 0.12.5: postprocessing no longer scales with how crowded your boxes are","comments":1,"category":"Announcements","created_at":"2026-08-04T11:36:30Z"},{"number":1415,"title":"Can we improve the memory usage of the slicing operation?","comments":2,"category":"Ideas","created_at":"2026-08-02T13:23:37Z"},{"number":1413,"title":"Proposal: Sparse CSR fast path for large Numba NMS and GreedyNMM workloads","comments":0,"category":"Ideas","created_at":"2026-08-01T15:58:13Z"},{"number":1397,"title":"Add support for Libreyolo","comments":2,"category":"Ideas","created_at":"2026-07-22T02:40:47Z"},{"number":1396,"title":"SAHI 0.12.2 is out: postprocessing no longer allocates a quadratic matrix","comments":0,"category":"Announcements","created_at":"2026-07-20T19:09:17Z"},{"number":1388,"title":"Inference for RF-DETRSeg Model","comments":1,"category":"Q&A","created_at":"2026-07-08T19:28:26Z"}],"latest_release":{"name":"0.14.0","tag":"0.14.0","published_at":"2026-11-01T01:43:00Z","body":"A minor release that moves the tests, docs and examples to YOLO26, defaults YOLOE to YOLOE-26, fixes bugs in the YOLOE, YOLO-World, torchvision, Hugging Face and mmdet wrappers, downloads and scripts, and makes COCO slicing and mask drawing faster with the same output.\n\n### 💥 Breaking Changes\n\n- **YOLOE loads `yoloe-26s-seg.pt` when no `model_path` is given** ([#1499](https://github.com/obss/sahi/pull/1499)) instead of `yoloe-11s-seg.pt`. Pass `model_path` to keep using YOLOE-11.\n- **The `hugging_face_universal_segmentation` model type is removed** ([#1498](https://github.com/obss/sahi/pull/1498)). It had no module behind it and always failed with `ModuleNotFoundError`. Use `huggingface_segmentation` for those models.\n\n### 🐛 Fixes\n\n- **YOLOE and YOLO-World predictions use the class names set with `set_classes`** ([#1499](https://github.com/obss/sahi/pull/1499)) instead of `'0'`, `'1'` and so on. A `category_mapping` you pass in is still kept.\n- **`model_type=\"yolo-world\"` loads** ([#1498](https://github.com/obss/sahi/pull/1498)). The registered class name was misspelled, so it always failed with `AttributeError`.\n- **The torchvision wrapper keeps every image of a batch** ([#1502](https://github.com/obss/sahi/pull/1502)) with its own shift amount and full shape, instead of only the last image.\n- **Hugging Face panoptic segmentation keeps batch outputs paired with their slices** ([#1502](https://github.com/obss/sahi/pull/1502)). An image with no segments was dropped, so the predictions after it got the wrong slice positions.\n- **mmdet inference no longer has a path that raises `NameError`** ([#1502](https://github.com/obss/sahi/pull/1502)).\n- **`python -m sahi.scripts.slice_coco` runs** ([#1501](https://github.com/obss/sahi/pull/1501)). It passed the builtin `slice` to Fire instead of `slicer`.\n- **`predict` adds up the export time of every image** ([#1501](https://github.com/obss/sahi/pull/1501)) instead of keeping only the last one, and no longer raises `KeyError` with `verbose=2` when there are no images.\n- **`predict_fiftyone` counts the prediction time of sliced predictions** ([#1501](https://github.com/obss/sahi/pull/1501)).\n- **`sahi env` shows pillow and every installed OpenCV distribution** ([#1501](https://github.com/obss/sahi/pull/1501)), and no longer lists tensorflow, which sahi does not use.\n- **Downloads write to a temporary file first** ([#1495](https://github.com/obss/sahi/pull/1495), [#1504](https://github.com/obss/sahi/pull/1504)), so an interrupted download never leaves a broken model file that later runs would load. The YOLOv5 helpers used by the YOLOv5 notebook now go through the same download code.\n\n### ⚡ Performance\n\nAll of these give the same output as before ([#1503](https://github.com/obss/sahi/pull/1503)).\n\n| Benchmark | v0.13.1 | v0.14.0 |\n|---|---|---|\n| `slice_coco` on 50 VisDrone images, 6493 sliced annotations | 1.48 s | 0.92 s |\n| Drawing 80 masks on a 4K image | 2.09 s | 0.15 s |\n| Drawing masks on 100 frames of the OpenCV `vtest.avi` video | 1.20 s | 0.12 s |\n\n- **Slicing computes each annotation's bbox and area once per image** instead of once per slice.\n- **Mask drawing blends only the area each mask covers** instead of the whole image.\n- **Sliced prediction and `SliceImageResult` indexing no longer rebuild the slice lists on every access**, and `predict` converts each image to an array once for all exports.\n\n### 🧹 Maintenance & CI\n\n- **Tests and CI commands use YOLO26** ([#1496](https://github.com/obss/sahi/pull/1496)) for detection, segmentation, OBB and ONNX export.\n- **CI installs CPU torch** ([#1494](https://github.com/obss/sahi/pull/1494)), which makes the CI caches about three times smaller.\n- **`from __future__ import annotations` is removed** ([#1500](https://github.com/obss/sahi/pull/1500)) now that Python 3.10 is the minimum, and ruff stops it from coming back.\n\n### 📚 Documentation\n\n- **The README, docs and examples use YOLO26** ([#1497](https://github.com/obss/sahi/pull/1497)) in","is_prerelease":false},"readme":true,"readme_zh":false,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}