Abstract:
One of the challenges facing the use of fisheye or hemispherical cameras in the task of
object recognition despite their wide field of view is the lack of a benchmark dataset
and the presence of distortions. Although there are a number of fisheye datasets avail able, they are usually focused on specific object recognition tasks. This is evident in
the case of KITTI-360 focused on autonomous driving and THEODORE which is fo cused on person recognition. In this work, Fisheye365, a dataset for generalized object
recognition is proposed. It contains 5.1 million images obtained by applying 8 differ ent transformations to the original Objects365 dataset. For testing purposes, COHI,
a benchmark fisheye testing dataset is modified to COHI-365. Data-centric methods
are applied to improve the performance of the YOLOv7 model on perspective and fish eye images. The YOLOv7 model trained on Fisheye365, YOLOv7-T2, outperforms
the YOLOv7 model trained only on Objects365, YOLOv7-0 by 4.8% when tested
on COHI-365. Finally, to further improve the performance of YOLOv7-T2, 16,000
synthetic images containing 13 iconic classes lying at the intersection of underrepre sented and underperforming classes of Objects365 were generated. YOLOv7-T2 was
finetuned on this synthetic dataset and had a least performance of 21% on 8 out of
the 13 classes