Ultralytics is a computer vision and AI toolkit for building cutting-edge YOLO series models, based on years of foundational research in computer vision and AI, and continuously updated to balance performance and flexibility. Its models are fast, accurate, and easy to use.
[Uses and Supported Tasks] Ultralytics YOLO excels at multiple tasks: object detection, instance segmentation, semantic segmentation, depth estimation, image classification, pose estimation, and oriented object detection (OBB), and can track detected objects across video frames.
[Model Versions] The project covers a variety of YOLO models from early YOLOv3 to the latest YOLO26. YOLO27 is still in the final stage of development, the model is not yet available, and there is no release date. The materials mention names such as YOLO27, YOLO26, YOLO11, and YOLOv8.
[Pretrained Models and Datasets] The YOLO26 series provides pretrained models: detection, segmentation, and pose estimation models are pretrained on COCO; semantic segmentation is pretrained on Cityscapes (19 classes); depth estimation is pretrained on a broad combination of multiple datasets and evaluated on NYU Depth V2, predicting per-pixel depth in meters; classification models are pretrained on ImageNet (1000 classes); OBB models are pretrained on DOTAv1 (15 classes). Tracking mode is compatible with detection, segmentation, pose, and OBB models. The listed pretrained models are automatically downloaded from the latest Ultralytics release on first use.
[Usage] Installation: install the ultralytics package in a Python>=3.8, PyTorch>=1.8 environment (including all dependencies). Other installation methods include Conda, Docker, and building from source via Git; see the quickstart guide for details. Command line: you can run it directly with the yolo command, which supports multiple tasks and modes and accepts additional arguments such as imgsz=640; see the YOLO CLI documentation for examples. Python: it can be integrated directly into Python projects, and accepts the same configuration arguments as the CLI; see the YOLO Python documentation for examples. The materials also provide some command examples for reproducing evaluation results, such as: yolo val detect data=coco.yaml device=0; yolo val segment data=coco.yaml device=0; yolo semantic val data=cityscapes.yaml device=0 imgsz=2048; yolo depth val model=yolo26n-depth.pt data=nyu-depth.yaml imgsz=768 device=0; yolo val classify data=path/to/ImageNet device=0; yolo val pose data=coco-pose.yaml device=0; yolo val obb data=DOTAv1.yaml device=0 split=test. Speed metrics are all reproduced based on batch=1, device=0|cpu, etc.; CPU speed uses ONNX export, GPU speed uses TensorRT export, and they are measured on instances such as Amazon EC2 P4d.
[Evaluation Metric Descriptions] For detection and segmentation, mAP val is the performance of a single model at a single scale on COCO val2017; semantic segmentation uses mIoU on the Cityscapes validation set; depth estimation uses the NYU Depth V2 Eigen test split, delta1 NYU represents the percentage of pixels whose predicted depth is within 1.25 times the ground truth (multiscale plus horizontal flip TTA with log least-squares alignment), abs_rel is the mean absolute relative error, rmse is the root mean square error in meters, and the speed metric is pure inference latency at imgsz=768, batch=1 (excluding pre-processing/post-processing); classification uses accuracy acc on the ImageNet validation set; pose estimation uses mAP val on COCO Keypoints val2017, focusing on the person class; OBB uses single-model multiscale mAP test on the DOTAv1 test set.
[Integration and Ecosystem] With leading AIntegrations with the I platform extend Ultralytics' functionality, covering tasks such as dataset annotation, training, visualization, and model management. Partners include Weights & Biases, Comet ML, Roboflow, and Intel OpenVINO; for more integrations, see Ultralytics Integrations. The project welcomes community contributions, provides a Contributing Guide, and has a survey to collect feedback.
[License and Support] Two types of licenses are offered: the AGPL-3.0 open-source license, suitable for students, researchers, and enthusiasts, encouraging open collaboration and knowledge sharing, with details in the LICENSE file; and the Ultralytics Enterprise License, for development and production scenarios, which allows seamless integration of Ultralytics software with AI models into business products and services (including internal tools, automated workflows, and production deployments), avoids the AGPL-3.0 open-source requirements, and requires contacting Ultralytics Licensing to obtain. Support and feedback channels include GitHub Issues (bug reports and feature requests), Discord, Reddit, and the Ultralytics community forum.
[Limitations and Missing Information] The YOLO27 model has not yet been released and has no release date; information such as pricing, specific commercial terms details, and platform coverage is not provided.
