DyMC-YOLO: How Dual-Stream Fusion Boosts Real-Time Detection to 46 FPS

Global, Source:

Object detection stands as a fundamental pillar of intelligent visual perception. It powers critical advancements in autonomous driving, security surveillance, and precision agriculture. Yet traditional visible-spectrum detectors often falter under adverse conditions like low illumination glare or heavy occlusion. Researchers have now introduced DyMC-YOLO a novel dual-stream detector that fuses RGB and infrared data to enhance performance across diverse environments.

DyMC-YOLO builds upon the YOLOv13 architecture chosen for its balance of accuracy and real-time efficiency. The system employs symmetric but independently parameterized backbone branches for RGB and infrared inputs. A key innovation is the Complementary Mask-Guided Feature Fusion module which performs channel-preserving spatial weighting at three semantic levels. Unlike previous methods that may double feature width or require complex projections this module generates a single complementary spatial mask to adaptively weight modalities based on image content.

This approach preserves direct feature-transmission paths through identity mappings in late-stage blocks. The model also introduces the Dynamic Multi-Path Detection Head. This component replaces fixed prediction structures with task-specific dynamic routing blocks. For each sample the head aggregates short-range medium-range and long-range prediction paths using sample-conditioned coefficients. This allows the network to dynamically adjust its receptive field for regression and classification tasks improving localization and context understanding without increasing computational latency significantly.

Evaluation on the M3FD traffic dataset and a self-constructed RGB-NIR grape dataset demonstrated superior accuracy-efficiency trade-offs. On M3FD the model achieved a mean Average Precision of 51.73% outperforming other lightweight YOLO-based fusion methods while maintaining a compact footprint of approximately 8 million parameters. On the agricultural grape dataset it reached 83.06% mAP showcasing robustness in cluttered low-light scenes.

The model exhibited strong resilience to cross-modal misregistration and missing modality scenarios retaining high absolute detection accuracy even under controlled perturbations. With an inference speed of nearly 46 FPS and low memory consumption DyMC-YOLO presents a viable solution for edge deployment on neural processing units. By effectively combining hierarchical feature fusion with dynamic prediction pathways this architecture sets a new standard for reliable real-time multimodal object detection in challenging visual conditions.