Generalized Object Detection on Fisheye Cameras for Autonomous Driving: Dataset, Representations and Baseline

Object detection is a comprehensively studied problem in autonomous driving.\nHowever, it has been relatively less explored in the case of fisheye cameras.\nThe standard bounding box fails in fisheye cameras due to the strong radial\ndistortion, particularly in the image's periphery. We explore better\nrepresentations like oriented bounding box, ellipse, and generic polygon for\nobject detection in fisheye images in this work. We use the IoU metric to\ncompare these representations using accurate instance segmentation ground\ntruth. We design a novel curved bounding box model that has optimal properties\nfor fisheye distortion models. We also design a curvature adaptive perimeter\nsampling method for obtaining polygon vertices, improving relative mAP score by\n4.9% compared to uniform sampling. Overall, the proposed polygon model improves\nmIoU relative accuracy by 40.3%. It is the first detailed study on object\ndetection on fisheye cameras for autonomous driving scenarios to the best of\nour knowledge. The dataset comprising of 10,000 images along with all the\nobject representations ground truth will be made public to encourage further\nresearch. We summarize our work in a short video with qualitative results at\nhttps://youtu.be/iLkOzvJpL-A.\n