Source-free Domain Adaptation for Video Object Detection Under Adverse Image Conditions

When deploying pre-trained video object detectors in real-world scenarios, the domain gap between training and testing data caused by adverse image conditions often leads to performance degradation. Addressing this issue becomes particularly challenging when only the pre-trained model and degraded videos are available. Although various source-free domain adaptation (SFDA) methods have been proposed for single-frame object detectors, SFDA for video object detection (VOD) remains unexplored. Moreover, most unsupervised domain adaptation works for object detection rely on two-stage detectors, while SFDA for one-stage detectors, which are more vulnerable to fine-tuning, is not well addressed in the literature. In this paper, we propose Spatial-Temporal Alternate Refinement with Mean Teacher (STAR-MT), a simple yet effective SFDA method for VOD. Specifically, we aim to improve the performance of the one-stage VOD method, YOLOV, under adverse image conditions, including noise, air turbulence, and haze. Extensive experiments on the ImageNetVOD dataset and its degraded versions demonstrate that our method consistently improves video object detection performance in challenging imaging conditions, showcasing its potential for real-world applications.

1. Introduction

Object detection in images and videos represents a pivotal task in computer vision, primarily owing to its extensive range of applications across diverse scenarios, such as intelligent surveillance systems and automated driving vehicles. Video object detection (VOD) aims to predict the bounding box and category information of all targeting objects in all video frames. Compared to single-frame object detection tasks, VOD enjoys the advantage of accessing additional information from the temporal dimension [34], [45], which often contains consistent semantics and multiple views of the same target to help enrich the feature space and facilitate superior performance.