A Dark Transformation-Equivariant Algorithm for Dark Object Detection

Tongxu Lin, Junyu Lin, Jingchao Wang, Zhengnan Deng, Guoheng Huang

Abstract: With the advent of deep learning, convolution neural networks (CNN) based object detection plays a critical role in fields such as autonomous driving. Unlike the normal-lit scenario, the dark environment can be challenging for object detection due to the low lighting and dark noise. To improve object detection in low-light conditions, we introduce a novel Dark Transformation-Equivariant (DTE) algorithm to explore feature consistency between the normal-lit and low-light domains. Specifically, on the one hand, we construct a dark transformation to simulate poor lighting conditions by darkening the regular images, considering the sensor noise. On the other hand, we capture the representation invariance of the dark transformation by encouraging feature consistency. Under the design of DTE, detectors are able to learn more discriminative representations of dark images for low-light object detection. Experiments on the ExDark dataset demonstrate the effectiveness of our method in improving detection performance when suffering from a dark environment and we achieve a 3.4% higher mAP than that of YOLOv3.