Tencent/YOLO-Master

What it solves

Real-time object detection (RTOD) models typically use static, dense computation, meaning they spend the same amount of processing power on a simple background as they do on a complex scene. This leads to computational redundancy and suboptimal performance. YOLO-Master introduces "compute-on-demand" to allow models to allocate computational resources based on the complexity of the input scene.

How it works

YOLO-Master integrates an Efficient Sparse Mixture-of-Experts (ES-MoE) block into the YOLO architecture. It uses a lightweight dynamic routing network to guide expert specialization during training and activates only the most relevant experts during inference. This allows the model to adaptively allocate FLOPs based on the input, reducing redundant computation while maintaining or improving precision.

Who it’s for

Developers and researchers working on real-time object detection who need to balance high accuracy with ultra-low latency, particularly for deployment on edge devices or in challenging, dense visual environments.

Highlights

• ES-MoE Integration: The first deep integration of Mixture-of-Experts into the YOLO architecture for general datasets.
• Performance Gains: Achieves 42.4% AP with 1.62ms latency on MS COCO, outperforming YOLOv13-N in both speed and accuracy.
• Parameter-Efficient Fine-Tuning: Includes architecture-agnostic LoRA support for various YOLO families (v3 through v12) and RT-DETR, significantly reducing training memory and storage.
• Advanced Tooling: Provides MoE pruning tools to remove redundant experts, Cluster-Weighted NMS (CW-NMS), and Sparse SAHI inference for accelerated small object detection.
• Extended Modules: Supports Mixture-of-Attention (MoA) and Mixture-of-Transformers (MoT) for multi-scale attention fusion and content-aware routing.