Domain Adaptive Object Detection for UAV-based Images by Robust Representation Learning and Multiple Pseudo-label Aggregation
Wu Ke, Jiaxin Chen, Miao Wang · 2024
Object detection on aerial images captured by Unmanned Aerial Vehicles (UAVs) has a wide range of applications. Due to the variations in illumination, weather conditions and scene backgrounds, the testing images (target domain) typically exhibit substantial discrepancies compared to training images (source domain). Considering the expensive annotation cost, it is crucial to develop domain adaptive object detection methods in the case of limited labeling resource in the target domain. However, most of existing approaches are designed for generic object detection without taking into account the unique characteristics of aerial images, leaving much room for improvement. To address this issue, in this paper we propose a novel method dubbed UAV-AdaptiveNet for domain adaptive object detection on UAV-based aerial images. Specifically, we present the cross-domain robust representation network (CDRN) by explicitly learning to disentangle the domain-invariant and domain-specific features. In the mean time, we develop the multiple pseudo-labels aggregation (MPA) for domain transfer learning based on the teacher-student framework, which further effectively mitigates the miss-detection on small-scale objects. Experimental results under various cross-domain settings and extensive ablation results clearly demonstrate the effectiveness of the proposed method, by comparing to the state-of-the-art approaches.