<p>The large domain gap between the normal-light datasets and low-light dataset lead to performance degradation on object detectors. Disentanglement-based unsupervised domain adaptation methods address this by separating domain-specific representations (DSR) from domain-invariant representations (DIR). However, existing methods focused exclusively on the separation purity of DIR but neglected its integrity and relied on source-target domain binary metrics for feature alignment, which limits detection accuracy and adaptability. To resolve these difficulties, we propose a low-light domain adaptation object detection framework. It adopts iterative residual feature fusion mechanism to separately enhance the two branches of DIR and DSR, improving their separability and ensuring purity and integrity. We further propose an auxiliary domain-guided triplet feature metric, which introduces additional feature references to facilitate cross-domain feature alignment. Experimental results demonstrate the effectiveness of the proposed method.</p>