Foreground Feature Selection and Alignment for Adaptive Object Detection

Abstract

Recently, remarkable progress has been witnessed in adaptive object detection, which aims to mitigate the distributional shifts between source domain and target domain. Domain-adversarial learning methods align the features of different levels to minimize the domain discrepancy, which have been proven effective for adapting object detectors. Most domain adaptation methods align whole-image features. Therefore, foreground alignment may be interfered by the backgrounds. In this work, we propose Foreground Feature Alignment Framework (FFAF) that strengthens the foreground alignment. One of our key contributions is the Foreground Selection Module (FSM), which captures the foreground features that are crucial for object detection and helpful for subsequent feature alignment. Additionally, we align the foreground features by integrating multi-level domain classifiers. Multi-level Domain adaptation (MDA) can simultaneously bridge the domain gap at various representation levels. We evaluate our method with multiple experiments, whose results demonstrate that our method achieves significant improvements in different cross-domain object detection tasks.