YOLO27: An Overview of Dual-Scale Architecture and Query-Based Real-Time Computer Vision

Abstract

This study presents a comprehensive overview of Ultralytics YOLO27 (also called as YOLOv27), examining its scale-adaptive architecture, detection mechanisms, preliminary performance, and implications for real-time computer vision. Unlike conventional YOLO families that largely scale common architectures across model sizes, YOLO27 introduces capacity-dependent architectural specialization. YOLO27n/s employ streamlined convolutional neural networks with strengthened high-resolution features and dual-scale prediction, supporting one-to-many detection with non-maximum suppression (NMS) and optional one-to-one NMS-free inference. Conversely, YOLO27m/l adopt query-based transformer decoding for native end-to-end NMS-free detection, while YOLO27l further integrates UltraViT with deep-stage self-attention for enhanced global-context modeling. Foreground alignment supervision reduces discrepancies between dense training and one-to-one inference. Preliminary COCO benchmarks report 42.3--60.4 mAP at 640-pixel resolution and 0.62--2.32 ms TensorRT 11 FP16 latency, while YOLO27l reaches 61.2 mAP at 800 pixels. Beyond object detection, this overview examines instance and semantic segmentation, depth estimation, classification, pose estimation, oriented detection, deployment trade-offs, current limitations, and future opportunities spanning multimodal vision, edge AI, autonomous systems, and robotic perception, providing a perspective on YOLO27's evolving role within real-time intelligent visual perception systems.

Keywords