Improving Cross-Domain Semi-Supervised Object Detection with Adversarial Domain Adaptation

Abstract

In autonomous driving, millions of frames with various scenarios for training deep object detectors is required. Labeling such a large number of frames is a costly process, therefore additional data sources support the training task. However, domain gaps from different cameras, weather, or locations typically limit the performance.We apply semi-supervised object detection, which leverages labeled source and pseudo-labeled target domain data in an iterative training paradigm. In addition, we newly include state-of-the-art adversarial style transfer into the semi-supervised training by stylizing images from source and target domains. This reduces the domain gap and improves pseudo-label quality in cross-domain semi-supervised training.In experiments and ablation studies, we show that our novel training framework can improve state-of-the-art detection performance by up to +10.1% on standard domain adaptation benchmarks.