Tencent/YOLO-Master

[CVPR2026]🚀🚀🚀Official code for the paper "YOLO-Master: MOE-Accelerated with Specialized Transformers for Enhanced Real-time Detection." *(YOLO = You Only Look Once)* 🔥🔥🔥

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" through a Mixture-of-Experts (MoE) architecture, allowing the model to adaptively allocate computational resources based on the complexity of the input image.

How it works

  • Efficient Sparse MoE (ES-MoE): The framework integrates MoE blocks into the YOLO architecture. Instead of a single dense layer, it uses multiple "experts" (specialized sub-networks).
  • Dynamic Routing: A lightweight routing network determines which experts are most relevant for a given input, activating only a subset of them during inference to reduce FLOPs.
  • MoE Optimization: It includes a load-balancing loss (MoELoss) to ensure experts are trained effectively and pruning tools to remove underutilized experts for further speedups.
  • PEFT Integration: It supports Parameter-Efficient Fine-Tuning (PEFT) via LoRA, allowing users to adapt models to new tasks using only a small fraction of trainable parameters.

Who it’s for

  • Computer Vision Engineers: Those building real-time detection systems that need to maximize the trade-off between accuracy and latency.
  • Edge Device Developers: Developers deploying models to hardware with limited compute budgets who need adaptive inference.
  • AI Researchers: Those exploring the intersection of MoE and vision tasks.

Highlights

  • Adaptive Computation: Transitions from static dense computation to input-adaptive allocation.
  • 高性能 (High Performance): YOLO-Master-N achieves 42.4% AP on MS COCO with 1.62ms latency, outperforming YOLOv13-N in both speed and accuracy.
  • Parameter-Efficient Fine-Tuning: Built-in LoRA support for the entire YOLO family (v3 through v12) and RT-DETR, reducing memory usage by up to 70%.
  • Comprehensive Tooling: Includes MoE pruning, diagnostic tools, and support for exporting to ONNX and TensorRT.

Related

  • Dispatch
    Mixture of Experts (MoE) Explained

    Mixture of Experts (MoE) allows transformer models to scale parameters while maintaining efficient pretraining and faster inference by activating only a subset of neural network experts per token.

  • Project
    lyuwenyu/RT-DETR

    A real-time object detection framework based on the Detection Transformer (DETR) architecture that aims to outperform YOLO models in speed and accuracy.

  • Project
    z1069614715/objectdetection_script

    A comprehensive toolkit for object detection that provides modified YOLO and RT-DETR architectures, model pruning and distillation scripts, and dataset processing tools for research and deployment.

  • Project
    yzfzzz/depth-detect

    A bicycle intelligent rear-view mirror system that fuses YOLO object detection with depth estimation to detect and alert users of fast-approaching dangerous vehicles.

  • Project
    Koldim2001/YOLO-Patch-Based-Inference

    A Python library that enables patch-based inference for YOLO models to improve the detection of small objects in large images through image tiling and result consolidation.