Improving Zero-shot Low-light Object Detection via Handling of Motion Blur

Tongxu Lin, Runqing Tang, fuhaozhang, Dong Chuyu, Dingguanzhi

Keywords: low-light object detection, zero-shot, motion blur

TL;DR: This paper collects a low-light dataset and propose an effective illumination-blur consistency framework for zero-shot low-light object detection.

Abstract: Zero-shot low-light object detection (ZLOD) presents great challenges as it aims to generalize detectors from the daytime domain to the nighttime domain without target data. Existing methods primarily focus on learning illumination consistency through daytime and synthetic nighttime image pairs, but they ignore a crucial characteristic that commonly co-exists with low illumination, i.e., motion blur. Nighttime images are particularly susceptible to motion-induced blur due to the long exposure times of cameras. Thus, solely considering illumination reduction without motion blur may be sub-optimal for ZLOD. To this end, we propose a novel Illumination-Blur Consistency (IBC) framework for ZLOD. Specifically, we synthesize nighttime images by considering illumination reduction and motion blur generation under a unified pipeline to access the complex nighttime domain. Then, we explore illumination-blur equivariant representations at the region and instance levels for better model adaptation. Consequently, IBC enables detectors to effectively generalize to the nighttime domain without relying on any dark data. Experimental results demonstrate the superior generalizability of our method. We also introduce a novel dataset named NightVision to expand the capacity of existing low-light benchmarks for community development.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 5978