Beyond bounding boxes: Segmentation supervision for robust object detection in fisheye images

Fisheye cameras pose significant object detection challenges due to severe radial distortion, rendering traditional axis-aligned bounding boxes suboptimal for warped object shapes. We propose a pipeline that transforms bounding box annotations into instance segmentation masks using the Segment Anything Model (SAM) and validate mask fidelity against expert ground truth in both rectilinear and distorted domains. We benchmark various models on the Fisheye8K dataset, demonstrating the architectural generalizability of our approach across YOLOv8, YOLOv11, and YOLOv12. Results show that segmentation-based supervision yields substantial performance gains, improving the mean average precision (mAP@[0.5:0.95]) by up to 10 absolute points over models trained with bounding boxes, and up to 12 points in distorted outer regions. Furthermore, our approach outperforms state-of-the-art methods and establishes a new benchmark for fisheye object detection. This work highlights the specific theoretical and empirical benefits of automated segmentation-based annotation within complex, distorted imaging domains.