Parameter Efficient Adaptation of Vision Large Language Models for Fisheye Object Detection

Résumé fourni par la source

Fisheye cameras are used for coverage in transportation systems and vehicles, yet distortions pose challenges to object detection algorithms. In this paper, Radial-Aware LoRA (RA-LoRA) is proposed, a distortion-aware parameter-efficient adaptation method that modulates low-rank updates based on radial position to address the spatially varying nature of fisheye distortion. RA-LoRA fine-tuning of Florence-2-large with only 0.5% additional parameters achieves a 181% relative improvement in mAP@50 over zero-shot inference, outperforming fully trained CNN detectors in fisheye imagery. Detection performance is assessed in all five object classes of the FishEye8K dataset (bike, bus, car, pedestrian, and truck). The results show that Florence-2-large with RA-LoRA achieves a mean average precision (mAP) of 0.724, outperforming YOLO26x (mAP: 0.581) and YOLO26 l (mAP: 0.605). Although the model requires approximately 4.5× more memory than YOLO26, the RA-LoRA adapters represent only 0.5% of the total parameters and can be merged into the base model at inference time with no additional overhead. This makes the proposed approach particularly suitable for server-side processing, offline traffic analysis, post-incident auditing, and high-accuracy surveillance applications where detection reliability outweighs the need for real-time embedded deployment.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.