Enhanced soft domain adaptation for object detection in the dark

Abstract

Unlike foggy conditions, domain adaptation is rarely facilitated in dark detection tasks due to the lack of dark datasets. We generate target low-light images via swapping the ring-shaped frequency spectrum of Exdark with Cityscapes, and surprisingly find the promotion is less satisfactory. The root lies in non-transferable alignment that excessively highlights dark backgrounds. To tackle this issue, we propose an Enhanced Soft Domain Adaptation (ESDA) framework to focus on background misalignment. Specifically, Soft Domain Adaptation (SDA) compensates for over-alignment of backgrounds by providing different soft labels for foreground and background samples. The Highlight Foreground (HF), by introducing center sampling, increases the number of high-quality background samples for training. Suppress Background (SB) weakens non-transferable background alignment by replacing foreground scores with backgrounds. Experimental results show SDA combined with HF and SB is sufficiently strengthened and achieves state-of-the-art performance using multiple cross-domain benchmarks. Note that ESDA yields 11.8% relative improvement on the real-world ExDark dataset.