YOLOv14:Unified Cross-Domain Real-Time Object Detectionwith Adaptive Multi-View Representation

2026-08-05

AI Summary

This work addresses the significant performance degradation of existing real-time object detectors under non-ideal inputs such as fisheye distortion, game-rendered imagery, aerial views, and 360° panoramas. To tackle complex geometric distortions and domain shifts, the authors propose a unified cross-domain real-time detection framework integrating four synergistic innovations: deformable region attention, Game2Real domain adaptation, multi-view conditional fusion, and adaptive augmentation. Key technical components include deformable attention mechanisms, AdaIN-based domain alignment, adversarial confusion, viewpoint embeddings, cross-view contrastive learning, dynamic scale routing, and scene-aware enhancement. The model achieves 49.1 mAP on COCO val2017 at 2.91 ms per image on an NVIDIA T4 GPU, while delivering substantial gains of 4.1, 6.6, 6.4, and 26.1 mAP on fisheye, panoramic, drone, and in-game character detection benchmarks, respectively.

Abstract

Real-time object detectors achieve remarkable accuracy under controlled conditions, yet degrade sharply on non-ideal inputs: fisheye distortion, game-rendered characters, aerial viewpoints, and 360° panoramas. We present YOLOv14, aunified detection framework addressing these challenges through four synergisticinnovations. (1) Deformable Area-Attention (D-AAttn) replaces rigid attentiongrids with learned 2D deformation fields, enabling adaptive sampling under geometric distortion. (2) Game2Real Domain Adaptation aligns rendered-game and photographic feature distributions via Adaptive Instance Normalization (AdaIN)and adversarial domain confusion, allowing game characters are detected as realhumans. (3) Multi-View Conditioning injects learned viewpoint embeddings intothe backbone with a cross-view contrastive loss that pulls same-class features fromdifferent perspectives closer. (4) An Adaptive Augmentation Policy automaticallyclassifies each input' scene type and routes to optimal augmentations, while a DynamicScaleRouter learns per-input feature pyramid weights. Together, YOLOv14achieves 49.1 mAP on COCO val2017 at 2.91 ms (T4 GPU), and delivers substantial gains on fisheye (+4.1 mAP), panorama (+6.6 mAP), drone (+6.4 mAP), andgame-character (+26.1 mAP) benchmarks

Problem

• real-time object detection
• cross-domain
• geometric distortion
• domain adaptation
• multi-view

Innovation

• Deformable Area-Attention
• Game2Real Domain Adaptation
• Multi-View Conditioning
• Adaptive Augmentation Policy
• Dynamic Scale Router