Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models (arxiv.org)

Ultralytics has unveiled YOLO26, a new family of unified real-time end-to-end vision models designed to enhance efficiency and accuracy in computer vision tasks. Unlike previous YOLO versions, YOLO26 eliminates traditional issues such as reliance on non-maximum suppression and the complexities of heavy detection heads, thereby streamlining the inference process. The model adopts a dual-head architecture for NMS-free operation while also introducing innovative training techniques, including a hybrid optimizer called MuSGD, Progressive Loss shifting supervision toward inference heads, and a new label assignment strategy (STAL) that ensures better coverage for small objects. This advancement is significant for the AI/ML community as it raises the bar for real-time vision systems, enabling improved performance across various applications such as detection, instance segmentation, pose estimation, and classification all within a single pipeline. YOLO26 operates efficiently across five different scales, achieving impressive metrics with a mean Average Precision (mAP) ranging from 40.9 to 57.5, while maintaining low latency of 1.7 to 11.8 ms on TensorRT. The introduction of YOLOE-26, which supports open-vocabulary inference, further expands its versatility, making it a substantial step forward in the realm of real-time computer vision solutions.