🚀 SAHI v0.12.0 Batch Inference, Torch-Free Core & New Open-Vocabulary Models

[onuralpszr]

SAHI v0.12.0 Batch Inference, Torch-Free Core & New Open-Vocabulary Models

Sliced Aided Hyper Inference detect the small stuff in big images.

This is one of the largest SAHI releases to date: 95 commits since 0.11.34 (rolling in the 0.11.35/0.11.36 hotfixes), spanning a re-architected post-processing engine, true batch inference, a lighter torch-free core, six new model families, and a full documentation overhaul.

---

🚀 Highlights

⚡ Batch inference, torch-free core & accelerated post-processing backends (#1336)

A ground-up reworking of how SAHI runs and merges predictions:

• Batch inference slices are now processed in batches end-to-end, dramatically improving throughput on GPU.
• Torch-free core the core slicing/ post-processing path no longer hard-depends on PyTorch. Install only what your model backend needs.
• Pluggable post-processing backends NMS / NMM now run on a selectable backend:
  • NumPy zero heavy deps, runs anywhere
  • Numba JIT-accelerated CPU path for large slice counts
  • TorchVision GPU-accelerated when torch is available
  • Backend auto-selects based on what's installed, and can be forced explicitly.

🎚️ Finer control over slicing & post-processing

• force_postprocess_type in get_sliced_prediction for explicit control over how overlapping detections are merged (feat: 🚀 add force_postprocess_type option to get_sliced_prediction for improved control over postprocessing #1346).
• Per-call confidence_threshold override across the prediction APIs tune confidence without rebuilding the model object (feat: 🚀 allow per-call confidence_threshold override in prediction APIs #1352).
• Progress bar + progress callback for get_sliced_prediction, available in both the Python API and CLI (Add progress bar and progress callback for get_sliced_prediction #1255).

🧠 New model support

• GroundingDINO (HuggingFace) zero-shot, text-prompted open-vocabulary detection. Describe what you're looking for in natural language and run it through SAHI's sliced pipeline perfect for finding small, rare, or unlabeled objects in large images ships with a dedicated demo notebook (Add GroundingDINO support through HuggingFace Transformers #1361).
• Universal segmentation from HuggingFace run any HF universal/panoptic segmentation model through SAHI's sliced pipeline (added support for universal segmentation models #1360).
• RF-DETR-Seg segmentation variant of RF-DETR (Add support for RF-DETR-Seg models #1315).
• YOLOE detection model (feat: ✨ add YOLOE detection model and update model type mapping #1268).
• YOLO-World open-vocabulary detection model (feat: ✨ add YOLOWorldDetectionModel and update model type mapping #1267).
• YOLO26 support across the Ultralytics backend, CLI, docs, and notebooks (fix: 🐞 update UltralyticsDetectionModel to support additional model formats and improve task handling #1321, fix: add yolo26 to ULTRALYTICS_MODEL_NAMES list #1322, feat: 🚀 ultralytics notebook updated to latest version of yolo26 #1356).

📚 Documentation overhaul

• Migrated docs to Zensical with full code typing & formatting cleanup (feat: 🚀 migrate to zensical docs and code typing & formatting #1344).
• Chinese (zh) translation of the documentation added and kept in sync (Docs: add Chinese translation for SAHI documentation #1253, docs: update Chinese docs to match latest English version #1332, docs(zh): sync Chinese translation with latest English version #1347).
• New API reference, ** post-processing backends guide**, security policy, and Code of Conduct (docs: ✨ add API reference documentation and improve argument descriptions in functions #1257, feat: 🚀 batch inference, torch-free core, and accelerated postprocessing backends #1336, docs: 📝 add security policy and reporting guidelines to enhance vulnerability management #1272, Add Contributor Covenant Code of Conduct #1349).

---

✨ Performance & Improvements

• Significantly faster post-processing — NMS, NMM, and GREEDYNMM now use a shapely STRtree spatial index, dramatically speeding up merging on images with many slices/detections (Significantly improve postprocessing speed (NMS, NMM, GREEDYNMM) by using STRtree from shapely #1248). Thanks @nikvo1!
• Faster read_image_as_pil for quicker slicing throughput (perf: ⚡️ enhance read_image_as_pil read speed for better slice speed  #1353).
• Improved performance & resource management in prediction and slicing (refactor: 🛠️ improve performance and resource management in prediction and slicing functions #1263).
• Better nms performance with correct handling of empty predictions (fix: 🐛 improve nms function for better performance and handle empty predictions #1288).
• Replaced pybboxes with a lightweight in-house yolo_bbox_to_voc_bbox (refactor: ♻️ replace pybboxes with custom yolo_bbox_to_voc_bbox function #1320).
• Dropped the pybboxes and pinned opencv-python version constraints for cleaner installs (remove opencv-python dependency version constraint #1325).

🐞 Bug Fixes

• Fixed empty bounding boxes caused by an empty shapely_annotation.multipolygon (Fix empty bbox because of empty shapely_annotation.multipolygon #1140).
• Fixed invalid segmentation masks for Detectron2 models (fix: Invalid Segmentation Mask for detectron2 models #1262).
• Corrected margin calculation in BoundingBox (fix: 🐛 correct margin calculation in BoundingBox class #1286).
• Fixed CHW-format image handling in read_image_as_pil (fix: 🐛 improve image handling in read_image_as_pil for CHW format #1287).
• Validate overlap ratios in get_slice_bboxes (must be < 1.0) (fix: 🐛 validate overlap ratios in get_slice_bboxes to ensure they are less than 1.0 #1285).
• Corrected error message for invalid model path in RTDetrDetectionModel (fix: 🐛 correct error message for invalid model path in RTDetrDetectionModel #1266).
• Fixed incorrect type annotations in the postprocess module (fix: 🐞 fix incorrect type annotations in postprocess module #1327).
• Ultralytics model now supports additional formats with improved task handling (fix: 🐞 update UltralyticsDetectionModel to support additional model formats and improve task handling #1321).
• Added pywinpty for Windows dev compatibility (fix: 🐞 add pywinpty dependency for Windows compatibility in dev and override-dependencies #1319).

🧹 Maintenance & CI

• Pinned all GitHub Actions to commit SHAs for supply-chain security (ci: update all of the github actions to sha for proper version and security #1351).
• Multi-OS CI matrix and clearer workflow naming (ci: 👷 update CI configuration for multi-OS support and improve naming #1334).
• Bumped to Python 3.12/3.13 in CI and docs (ci: 👷 github actions updates and doc python version update to 3.13 #1259, ci: 🛠️ update Python version to 3.12 and upgrade setup-uv action to v7.0.0 #1260).
• Removed deprecated YOLOv5 helpers, legacy requirements.txt, MMDet workflow, and unused Netlify config (refactor: ♻️ remove deprecated yolov5 functions and serialize methods #1326, chore: delete requirements.txt #1342, chore: delete .github/workflows/mmdet.yml #1341, chore: 🧹 remove unused netlify file #1335).
• numpy<3.0, torchvision 0.23.0, and many dependency bumps via Dependabot (now also covering pip).

---

🙏 Contributors

Thanks to everyone who contributed to this release:
@onuralpszr, @fcakyon, @siromermer, @ZephyrKeXiner, @yogendrasinghx, @srikrishnavignesh, @ibuldakov, @ducviet00, @volks73, @RizwanMunawar, @nikvo1, @vinnik-dmitry07, and @gboeer and @dependabot for keeping dependencies fresh.

🌟 New Contributors

• @vinnik-dmitry07 made their first contribution in Fix empty bbox because of empty shapely_annotation.multipolygon #1140
• @nikvo1 made their first contribution in Significantly improve postprocessing speed (NMS, NMM, GREEDYNMM) by using STRtree from shapely #1248
• @volks73 made their first contribution in fix: Invalid Segmentation Mask for detectron2 models #1262
• @ZephyrKeXiner made their first contribution in Docs: add Chinese translation for SAHI documentation #1253
• @yogendrasinghx made their first contribution in refactor: ♻️ remove deprecated yolov5 functions and serialize methods #1326
• @ibuldakov made their first contribution in Add support for RF-DETR-Seg models #1315
• @srikrishnavignesh made their first contribution in added support for universal segmentation models #1360
• @siromermer made their first contribution in Add GroundingDINO support through HuggingFace Transformers #1361

Full Changelog: 0.11.34...0.12.0

---

This discussion was created from the release 🚀 SAHI v0.12.0 — Batch Inference, Torch-Free Core & New Open-Vocabulary Models.

[gboeer]

Awesome update! Thanks for all the work everyone.