Fisheye image augmentation for overcoming domain gaps with the limited dataset

Hyeseong Lee, Sunmin Park, Miyoung Lee; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 3085-3093

The camera sensor plays a pivotal role in the perception systems of self-driving cars. Among various lens types, the fisheye lens provides an ultra-wide field of view, allowing vehicle designers to achieve a full 360-degree view with fewer cameras, thereby reducing production costs. However, annotated fisheye image datasets, particularly those with semantic labels, are significantly scarcer than those for conventional cameras. Additionally, collecting and labeling large-scale datasets for training deep learning models is time-consuming and expensive. Lastly, an existing method, which applies radial distortion to images has a domain gap problem stemming from their narrower FoV. To address this challenge, we propose a novel data augmentation algorithm that leverages both camera images and LiDAR point clouds to augment training data for domain adaptation. To validate the effectiveness of our approach, we employ convolutional neural networks and vision transformers, demonstrating that our augmentation method improves model performance, especially in detecting small objects. Our findings highlight the potential of LiDAR-assisted augmentation in mitigating data scarcity issues and domain gaps in fisheye-based perception systems for autonomous driving. The code will be opened as the acceptance of the paper.