SAMO: Semantic-Aware Model Optimization for Source-Free Domain Adaptive Object Detection

Wenzhang Zhou, Caiyu Zhang, Baojie Fan, Xiangzhu Meng, Dongyin Liu, Lianlei Shan

Keywords: Source-Free Domain Adaptive, Object Detection, Semantic-Aware, Model Optimization

Abstract: Source-Free Adaptive (SFA) object detection aims to adapt a pre-trained detector from a labeled source domain to an unlabeled target domain without access to source data. To address the performance degradation commonly observed in this setting, we introduce a novel Semantic-Aware Model Optimization (SAMO) framework that explicitly models the interplay between semantic perception and model optimization. Specifically, SAMO first performs semantic perception by categorizing target-domain regions into a confidence spectrum (from high to low) based on pseudo-labels generated by a static base detector. Building on this perception, we design an optimization strategy via detector decomposition and disentangled training. The detector is decoupled into shared modules and a specialized expert group, the latter equipped with a semantically aware gating network. Through disentangled training, expert units and shared modules are independently optimized according to their respective confidence spectrum ranges, thereby maintaining discriminative capacity across diverse semantic levels. Meanwhile, the gating network adaptively integrates expert features through learnable attention weights, enabling dynamic discrimination across semantic categories. This framework effectively alleviates the source-domain performance degradation caused by low-confidence regions, while simultaneously achieving significant improvements in the target domain. Extensive experiments on three domain adaptation benchmarks demonstrate the superior generalization ability of SAMO. Our code will be released.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 2704