1. DAViMNet: SSMs-Based Domain Adaptive Object Detection
- Author
-
Doruk, A. Enes and Ates, Hasan F.
- Subjects
Computer Science - Computer Vision and Pattern Recognition - Abstract
Unsupervised domain adaptation (UDA) for object detection adapts models trained on labeled source domains to unlabeled target domains, ensuring robust performance across domain shifts. Transformer-based architectures excel at capturing long-range dependencies but face efficiency challenges due to their quadratic attention complexity, which limits scalability in UDA tasks. To address these issues, we propose a hybrid domain-adaptive Mamba Transformer architecture that combines Mamba's efficient state-space modeling with attention mechanisms to tackle domain-specific spatial and channel-wise variations. Each hybrid block integrates domain-adaptive Mamba blocks and attention mechanisms: Domain-Adaptive Mamba employs spatial and channel state-space models to adaptively model domain variations, while attention mechanisms leverage self-attention for intra-domain feature enhancement and cross-attention for effective source-target alignment. Our approach processes both shallow and deeper features, employing an entropy-based knowledge distillation framework with margin ReLU to emphasize discriminative features and suppress noise. Gradient Reversal Layers enable adversarial alignment across network layers, while entropy-driven gating attention with random perturbations refines target features and mitigates overfitting. By unifying these components, our architecture achieves state-of-the-art performance in UDA object detection, balancing efficiency with robust generalization.
- Published
- 2025