SAHI: 切片辅助高效推理

一个轻量级的计算机视觉库,可实现大规模的目标检测和实例分割

teaser

downloads downloads License pypi version conda version Continuous Integration
Context7 MCP llms.txt ci arXiv Open In Colab HuggingFace Spaces Sliced/tiled inference DeepWiki built-with-material-for-mkdocs
##
概览
SAHI 通过启用**切片推理**来检测大图像中的小物体,从而帮助开发人员克服了对象检测中的实际挑战。它支持各种流行的检测模型并提供易于使用的 API。
🌐 [English](../../README.md) | 🇨🇳 [简体中文](README.md)
| 命令 | 描述 | |---|---| | [predict](https://github.com/obss/sahi/blob/main/docs/cli.md#predict-command-usage) | 使用任意 [ultralytics](https://github.com/ultralytics/ultralytics)/[mmdet](https://github.com/open-mmlab/mmdetection)/[huggingface](https://huggingface.co/models?pipeline_tag=object-detection&sort=downloads)/[torchvision](https://pytorch.org/vision/stable/models.html#object-detection) 模型进行切片或标准视频 / 图像预测 - 参见 [命令行指南](cli.md#predict-command-usage) | | [predict-fiftyone](https://github.com/obss/sahi/blob/main/docs/cli.md#predict-fiftyone-command-usage) | 使用任意支持的模型进行切片或标准预测,并在 [fiftyone应用](https://github.com/voxel51/fiftyone) 中探索结果 - [了解更多](fiftyone.md) | | [coco slice](https://github.com/obss/sahi/blob/main/docs/cli.md#coco-slice-command-usage) | 自动切片 COCO 标注和图像文件 - 参见 [切片工具](slicing.md) | | [coco fiftyone](https://github.com/obss/sahi/blob/main/docs/cli.md#coco-fiftyone-command-usage) | 在 [fiftyone ui](https://github.com/voxel51/fiftyone) 中探索 COCO 数据集的多个预测结果,按错误检测数量排序 | | [coco evaluate](https://github.com/obss/sahi/blob/main/docs/cli.md#coco-evaluate-command-usage) | 针对给定的预测和真实数据评估 COCO 的类级别 AP 和 AR - 查看 [COCO 工具](coco.md) | | [coco analyse](https://github.com/obss/sahi/blob/main/docs/cli.md#coco-analyse-command-usage) | 计算并导出多种错误分析图表 - 参见 [完整指南](../README.md) | | [coco yolo](https://github.com/obss/sahi/blob/main/docs/cli.md#coco-yolo-command-usage) | 将任意 COCO 数据集自动转换为 [ultralytics](https://github.com/ultralytics/ultralytics) 格式 | ### 社区认可 [📜 引用 SAHI 的出版物列表(当前超过 600 篇)](https://scholar.google.com/scholar?hl=en&as_sdt=2005&sciodt=0,5&cites=14065474760484865747&scipsc=&q=&scisbd=1) [🏆 使用 SAHI 的竞赛获奖者列表](https://github.com/obss/sahi/discussions/688) ### AI 工具认可 SAHI 的文档已在 [Context7 MCP](https://context7.com/obss/sahi) 中建立索引,为 AI 编码助手提供最新的,版本特定的代码示例和 API 参考。我们还提供了一个遵循 AI 可读文档新兴标准的 [llms.txt](https://context7.com/obss/sahi/llms.txt) 文件。要将 SAHI 文档集成到您的 AI 开发工作流程中,请查看 [Context7 MCP 安装指南](https://github.com/upstash/context7#%EF%B8%8F-installation). ##
安装
### 基本安装 ```bash pip install sahi ```
详细安装说明(点击展开) - 安装您所需的 PyTorch 和 torchvision 版本: ```console pip install torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/cu126 ``` (为了获得mmdet框架的支持,您需要安装torch 2.1.2版本): ```console pip install torch==2.1.2 torchvision==0.16.2 --index-url https://download.pytorch.org/whl/cu121 ``` - 安装您所需的检测框架 (ultralytics): ```console pip install ultralytics>=8.3.161 ``` - 安装您所需的检测框架 (huggingface): ```console pip install transformers>=4.49.0 timm ``` - 安装您所需的检测框架 (yolov5): ```console pip install yolov5==7.0.14 sahi==0.11.21 ``` - 安装您所需的检测框架 (mmdet): ```console pip install mim mim install mmdet==3.3.0 ``` - 安装您所需的检测框架 (roboflow): ```console pip install inference>=0.51.5 rfdetr>=1.6.2 ```
##
快速开始
### 教程 - [SAHI 简介](https://medium.com/codable/sahi-a-vision-library-for-performing-sliced-inference-on-large-images-small-objects-c8b086af3b80) - 请查阅 [完整的文档](../README.md) 以了解高级用法。 - [官方论文](https://ieeexplore.ieee.org/document/9897990) (ICIP 2022 oral) - [预训练权重 和 ICIP 2022 论文文件](https://github.com/fcakyon/small-object-detection-benchmark) - [视频教程(2025年)](https://www.youtube.com/watch?v=ILqMBah5ZvI) (推荐) - [使用 FiftyOne 可视化并评估 SAHI 的预测结果](https://voxel51.com/blog/how-to-detect-small-objects/) - [《探索 SAHI》——来自 learnopencv.com 的研究文章](https://learnopencv.com/slicing-aided-hyper-inference/) - [Encord 对 Slicing Aided Hyper Inference(SAHI)的解读](https://encord.com/blog/slicing-aided-hyper-inference-explained/) - [视频教程:SAHI 在小目标检测中的应用](https://www.youtube.com/watch?v=UuOJKxn-M8&t=270s) - [视频推理支持现已上线](https://github.com/obss/sahi/discussions/626) - [Kaggle notebook](https://www.kaggle.com/remekkinas/sahi-slicing-aided-hyper-inference-yv5-and-yx) - [卫星图像目标检测](https://blog.ml6.eu/how-to-detect-small-objects-in-very-large-images-70234bab0f98) - [误差分析绘图 & 评估](https://github.com/obss/sahi/discussions/622) (推荐) - [交互式结果可视化与检查](https://github.com/obss/sahi/discussions/624) (推荐) - [COCO 数据集转换](https://medium.com/codable/convert-any-dataset-to-coco-object-detection-format-with-sahi-95349e1fe2b7) - [切片操作 notebook 示例](../../demo/slicing.ipynb) - `YOLOX` + `SAHI` 示例: sahi-yolox - `YOLO12` + `SAHI` 实战教程: sahi-yolo12 - `YOLO11-OBB` + `SAHI` 实战教程: sahi-yolo11-obb (NEW) - `YOLO11` + `SAHI` 实战教程: sahi-yolo11 - `Roboflow/RF-DETR` + `SAHI` 实战教程: roboflow (NEW) - `RT-DETR v2` + `SAHI` 实战教程: sahi-rtdetrv2 (NEW) - `RT-DETR` + `SAHI` 实战教程: sahi-rtdetr - `HuggingFace` + `SAHI` 实战教程: sahi-huggingface - `YOLOv5` + `SAHI` 实战教程: sahi-yolov5 - `MMDetection` + `SAHI` 实战教程: sahi-mmdetection - `TorchVision` + `SAHI` 实战教程: sahi-torchvision sahi-yolox ### 与框架无关的切片/标准预测 sahi-predict 请在 [CLI 文档](cli.md#predict-command-usage) 中查找关于使用 `sahi predict` 命令的详细信息,并查阅 [预测 API](predict.md) 以了解高级用法。 请在 [视频推理教程](https://github.com/obss/sahi/discussions/626) 中查找关于视频推理的详细信息。 ### 误差分析绘图 & 评估 sahi-analyse 请在 [误差分析绘图 & 评估](https://github.com/obss/sahi/discussions/622) 中查找相关的详细信息。 ### 交互式结果可视化与检查 sahi-fiftyone 探索 [FiftyOne 集成](fiftyone.md) 以实现交互式可视化与检查。 ### 其他实用工具 请查阅 [完整的 COCO 工具指南](coco.md) 了解 YOLO 格式转换、数据集切片、子采样、筛选、合并与分割等操作。了解更多关于 [切片工具](slicing.md) ,以实现对图像和数据集切片参数的精细控制。 ##
引用
如果您在您的工作中使用了这个包,请如下文引用: ```bibtex @article{akyon2022sahi, title={Slicing Aided Hyper Inference and Fine-tuning for Small Object Detection}, author={Akyon, Fatih Cagatay and Altinuc, Sinan Onur and Temizel, Alptekin}, journal={2022 IEEE International Conference on Image Processing (ICIP)}, doi={10.1109/ICIP46576.2022.9897990}, pages={966-970}, year={2022} } ``` ```bibtex @software{obss2021sahi, author = {Akyon, Fatih Cagatay and Cengiz, Cemil and Altinuc, Sinan Onur and Cavusoglu, Devrim and Sahin, Kadir and Eryuksel, Ogulcan}, title = {{SAHI: A lightweight vision library for performing large scale object detection and instance segmentation}}, month = nov, year = 2021, publisher = {Zenodo}, doi = {10.5281/zenodo.5718950}, url = {https://doi.org/10.5281/zenodo.5718950} } ``` ##
贡献者
欢迎贡献!请参阅我们的 [贡献指南](../../CONTRIBUTING.md) 来开始使用. 感谢所有贡献者🙏!