CDMA-VP: A Cross-Domain Multi-Agent Visual Perception Engine with Diffusion-Based Adaptation
Year: 2026
Type: conference-paper
Abstract: Visual perception systems face critical challenges when deployed across diverse scenarios due to domain shift from varying imaging conditions, device configurations, and data distributions. This paper proposes CDMA-VP, a cross-domain multi-agent visual perception engine that addresses cross-domain adaptation through three key components: (1) six heterogeneous visual agents for diverse feature extraction, (2) a diffusion-based domain adaptation module for feature distribution alignment, and (3) an adaptive fusion mechanism based on Chair-Varshney optimal fusion theory. The multi-agent architecture integrates ResNet-50, FPN, MobileNetV3, DenseNet-121, EfficientNet-B3, and Vision Transformer to capture complementary visual features. The diffusion-based adaptation module progressively aligns feature distributions through iterative denoising, effectively mitigating domain shift. The adaptive fusion mechanism dynamically combines multi-agent predictions based on confidence scores and historical performance. On standard benchmarks (GTA5→Cityscapes, SYNTHIA→Cityscapes), CDMA-VP achieves 72.5% and 64.8% mIoU, improving over DAFormer by 4.2% and 3.9%, respectively. The adaptive agent selection mechanism reduces computational overhead by 50.0% while maintaining near-optimal performance with only 0.6% mIoU degradation.
Authors: Jianmin Wang, Jiahui Cao, Tianzuo Yuan
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7208765778