Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a new benchmark, FishBEVOD, for evaluating BEV object detection using mixed pinhole and fisheye camera setups, addressing the performance degradation of existing models designed primarily for pinhole cameras. They convert the KITTI-360 dataset into nuScenes format and evaluate three adaptations: rectification, distortion-aware view transformation modules (VTMs), and polar coordinate representations. The results show that projection-free architectures are more robust to fisheye distortion, providing practical guidelines for designing robust 3D perception systems.
Fisheye lenses can actually *improve* BEV object detection, but only if you ditch pinhole-centric architectures for projection-free approaches.
Modern autonomous driving systems increasingly rely on mixed camera configurations with pinhole and fisheye cameras for full view perception. However, Bird's-Eye View (BEV) 3D object detection models are predominantly designed for pinhole cameras, leading to performance degradation under fisheye distortion. To bridge this gap, we introduce a multi-view BEV detection benchmark with mixed cameras by converting KITTI-360 into nuScenes format. Our study encompasses three adaptations: rectification for zero-shot evaluation and fine-tuning of nuScenes-trained models, distortion-aware view transformation modules (VTMs) via the MEI camera model, and polar coordinate representations to better align with radial distortion. We systematically evaluate three representative BEV architectures, BEVFormer, BEVDet and PETR, across these strategies. We demonstrate that projection-free architectures are inherently more robust and effective against fisheye distortion than other VTMs. This work establishes the first real-data 3D detection benchmark with fisheye and pinhole images and provides systematic adaptation and practical guidelines for designing robust and cost-effective 3D perception systems. The code is available at https://github.com/CesarLiu/FishBEVOD.git.