Sampling-Free Universal Domain Adaptive Object Detection with Synergistic Optimization Alignment

Abstract

Universal Domain Adaptive Object Detection (UniDAOD) addresses domain shifts in open-world scenarios without assuming pre-defined shared categories between source and target domains. Despite progress in this field, existing methods rely heavily on thresholding techniques to filter private category samples, which introduces challenges in hyperparameter tuning and handling noisy data. To overcome these limitations, we propose Synergistic Optimization Alignment (SOA), a sampling-free framework that tackles two challenges inherent in UniDAOD: multi-scale discriminator discrepancy and task-dominant gradient conflict. To address the discrepancy among multi-scale discriminators, we propose Multi-Scale Discriminator Alignment (MDA) to adaptively aggregate under-aligned discriminators for shared classes, ensuring consistent feature alignment across scales. To address task-dominant gradient conflicts, we propose Gradient Coordinated Optimization (GCO), which suppresses conflicting domain-alignment gradients and coordinates their magnitudes to prevent domain over-alignment, thereby mitigating private-class interference while preserving beneficial domain-invariant supervision. Extensive experiments on open-set, partial-set, and closed-set domain adaptation benchmarks demonstrate that SOA outperforms state-of-the-art UniDAOD and DAOD methods. Code is available at https://github.com/zyfone/SOA.