PaperlayerView exact revision on arXiv

Source-linked AI summary

YOLO Meets Mixture-of-Experts: Adaptive Expert Routing for Robust Object Detection

Ori Meiraz, Sharon Shalev, Avishai Weizman

arXiv:2511.13344v4cs.CV

TL;DR

Object detection is a fundamental, challenging task of identifying and localizing objects in images or video. The paper introduces adaptive routing among YOLOv9-T experts and reports mAP and AR gains over other YOLOv9-T models on COCO and VisDrone.

  • Problem

    Object detection remains a fundamental, challenging task of identifying and localizing objects in images or video frames.

    Source: §I · p. 1

  • Method

    The paper uses a router to adaptively assign responsibility among multiple YOLOv9-T experts, enabling specialization in feature-map space.

    Source: §IV · p. 2, §III · p. 2

  • Results

    The model outperformed other YOLOv9-T models on COCO and VisDrone, with consistent mAP and AR gains across separate- and combined-dataset training.

    Source: §III · p. 2

  • Takeaways & Limitations

    The authors conclude that this design improves detection performance and robustness relative to a single model, with potential demonstrated on COCO and VisDrone.

    Source: §IV · p. 2

Abstract

from arXiv · show

This paper presents a novel Mixture-of-Experts framework for object detection, incorporating adaptive routing among multiple YOLOv9-T experts to enable dynamic feature specialization and achieve higher mean Average Precision (mAP) and Average Recall (AR) compared to a single YOLOv9-T model.

I. INTRODUCTION

The introduction frames object detection as a fundamental, challenging task and introduces a YOLOv9-T-based Mixture-of-Experts architecture.

I. INTRODUCTION

  • Object detection identifies and localizes objects in images or video frames and is described as fundamental and challenging.Source: §I · p. 1
  • CNN-based detectors including R-CNN, SSD, and YOLO – improved accuracy and speed; YOLOv9 outperformed other YOLO-series models [7].Source: §I · p. 1
  • The MoE paradigm [9] uses specialized experts and dynamic routing, combining outputs as weighted mixtures or hard expert selections.Source: §I · p. 1
  • For logit outputs, the fused output z combines expert logits z_e using router weights α_e.Source: §I · p. 1
  • The paper introduces a Mixture-of-Experts architecture based on YOLOv9-T.Source: §I · p. 1

ORCID: 0009-0004-1182-8601

The passage describes YOLOv9 as the basis for an architecture designed to improve detection performance and robustness.

ORCID: 0009-0004-1182-8601

II. PROPOSED MOE-BASED YOLOV9 ARCHITECTURE

The architecture applies Mixture-of-Experts routing at YOLOv9’s three feature-map scales, using expert features to generate adaptive weights that combine class and bounding-box predictions. A load-balancing term encourages uniform expert selection alongside the standard detection loss.

II. PROPOSED MOE-BASED YOLOV9 ARCHITECTURE

  • YOLOv9 processes each image at three feature-map resolutions, with bounding-box distances discretized into bins for probabilistic fusion across experts.The three resolutions are 8, 16, and 32; each level detects objects at a corresponding scale.Source: §II · p. 1
  • Following Enhanced MoE, routing uses intermediate features from individual experts rather than the raw image, with expert features fused at each level.Each expert extracts features from a layer preceding its output head.Source: §II · p. 1
  • At each level, the router input combines expert outputs and their concatenated representations through reweighted Hadamard fusion.The fusion captures cross-expert interactions and uses a learnable weighting parameter.Source: §II · pp. 1–2
  • A lightweight CNN with downsampling and a fully connected layer processes fused features to produce expert routing weights, which softmax normalizes and uses to adaptively modulate classification and bounding-box prediction outputs.Source: §II · p. 2
  • The total loss combines standard YOLOv9 detection loss with a load-balancing term that encourages uniform expert selection.The balancing term uses each expert’s routed-sample fraction and mean routing probability at each resolution.Source: §II · p. 2

III. EXPERIMENTS AND RESULTS

Across COCO and VisDrone, the two-expert YOLOv9-T MoE outperformed YOLOv9-T variants, with mAP and AR gains under separate- and combined-dataset training.

  • Experiments used COCO and VisDrone, evaluating Person, Vehicle, Bicycle, and Motorcycle after 50 epochs with YOLOv9 default hyperparameters.Source: §III · p. 2
  • Gains in mAP and AR persist for separately trained datasets and a single model trained on combined sets, outperforming other YOLOv9-T models on both datasets.Source: Table I · p. 2, §III · p. 2
  • The tested model used two YOLOv9-T experts (λlb = 0.5), dataset-pretrained expert weights, and routing modules trained from scratch.Source: §III · p. 2

IV. CONCLUSIONS AND FUTURE WORK

The paper concludes that adaptive routing among YOLOv9-T experts supports feature-map specialization and improves detection performance and robustness over a single model. Future work will explore broader model sizes, routing efficiency, video, and multimodal inputs.

  • The framework combines MoE with YOLOv9 adaptive routing between YOLOv9-T experts and improves detection performance and robustness compared with a single model.Source: §IV · p. 2
  • Future work will test other YOLO variants and sizes, more efficient routing, temporal video data, and multimodal inputs.Source: §IV · p. 2
Loading 2511.13344v4…