Camera-wise Training for Enhanced Omni-directional 2D Object Detection

Hyung-Joon Jeon, Duong Nguyen-Ngoc Tran, Long Hoang Pham, Huy-Hung Nguyen, Tai Huu-Phuong Tran, Jae Wook Jeon

Abstract: In this paper, we propose a method to perform training and inference with multiple instances of the same deep neural network architecture on images taken from cameras of different directions. Across multiple cameras, depending on each of their directional characteristics, objects viewed from the cameras can form slightly different distributions in visual features. Regarding this, we emphasize the importance of camera-wise training on multiple instances of a given deep neural network for object detection. Given the Waymo Open Perception Dataset, we used multiple instances of the YOLOv5x6 architecture and trained each of them per camera. Such a training scheme on the Training Set achieves better training progression, and the inference results are shown to have AP/L1 as high as 0.6679 on the Testing Set.