Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Real-Time Object Detection Meets DINOv3

Shihua Huang, Yongjie Hou, Longfei Liu, Xuanlong Yu, Xi Shen

TL;DR

DEIMv2 tackles real-time object detection across a broad spectrum of deployment budgets by integrating the strong semantics of DINOv3 with a lightweight Spatial Tuning Adapter (STA) and HGNetv2 pruning for ultra-light models. The method combines an RT-DETR–style architecture, ViT-based backbones for larger variants, and HGNetv2 backbones for ultra-light models, augmented by an efficient decoder and an enhanced Dense O2O with Copy-Blend. Key contributions include the STA for multi-scale feature extraction without costly parameters, a streamlined decoder, and the Dense O2O augmentation, yielding state-of-the-art COCO AP across S/M/L/X and competitive ultra-light performance (e.g., S > 50 AP with under 10M params; Pico achieving 38.5 AP with 1.5M params). The results demonstrate a unified, scalable design that maintains high accuracy while reducing parameters and FLOPs, enabling deployment from edge devices to GPU servers; future work may further boost small-object detection and latency optimization.

Abstract

Driven by the simple and effective Dense O2O, DEIM demonstrates faster convergence and enhanced performance. In this work, we extend it with DINOv3 features, resulting in DEIMv2. DEIMv2 spans eight model sizes from X to Atto, covering GPU, edge, and mobile deployment. For the X, L, M, and S variants, we adopt DINOv3-pretrained or distilled backbones and introduce a Spatial Tuning Adapter (STA), which efficiently converts DINOv3's single-scale output into multi-scale features and complements strong semantics with fine-grained details to enhance detection. For ultra-lightweight models (Nano, Pico, Femto, and Atto), we employ HGNetv2 with depth and width pruning to meet strict resource budgets. Together with a simplified decoder and an upgraded Dense O2O, this unified design enables DEIMv2 to achieve a superior performance-cost trade-off across diverse scenarios, establishing new state-of-the-art results. Notably, our largest model, DEIMv2-X, achieves 57.8 AP with only 50.3 million parameters, surpassing prior X-scale models that require over 60 million parameters for just 56.5 AP. On the compact side, DEIMv2-S is the first sub-10 million model (9.71 million) to exceed the 50 AP milestone on COCO, reaching 50.9 AP. Even the ultra-lightweight DEIMv2-Pico, with just 1.5 million parameters, delivers 38.5 AP, matching YOLOv10-Nano (2.3 million) with around 50 percent fewer parameters. Our code and pre-trained models are available at https://github.com/Intellindust-AI-Lab/DEIMv2

Real-Time Object Detection Meets DINOv3

TL;DR

DEIMv2 tackles real-time object detection across a broad spectrum of deployment budgets by integrating the strong semantics of DINOv3 with a lightweight Spatial Tuning Adapter (STA) and HGNetv2 pruning for ultra-light models. The method combines an RT-DETR–style architecture, ViT-based backbones for larger variants, and HGNetv2 backbones for ultra-light models, augmented by an efficient decoder and an enhanced Dense O2O with Copy-Blend. Key contributions include the STA for multi-scale feature extraction without costly parameters, a streamlined decoder, and the Dense O2O augmentation, yielding state-of-the-art COCO AP across S/M/L/X and competitive ultra-light performance (e.g., S > 50 AP with under 10M params; Pico achieving 38.5 AP with 1.5M params). The results demonstrate a unified, scalable design that maintains high accuracy while reducing parameters and FLOPs, enabling deployment from edge devices to GPU servers; future work may further boost small-object detection and latency optimization.

Abstract

Driven by the simple and effective Dense O2O, DEIM demonstrates faster convergence and enhanced performance. In this work, we extend it with DINOv3 features, resulting in DEIMv2. DEIMv2 spans eight model sizes from X to Atto, covering GPU, edge, and mobile deployment. For the X, L, M, and S variants, we adopt DINOv3-pretrained or distilled backbones and introduce a Spatial Tuning Adapter (STA), which efficiently converts DINOv3's single-scale output into multi-scale features and complements strong semantics with fine-grained details to enhance detection. For ultra-lightweight models (Nano, Pico, Femto, and Atto), we employ HGNetv2 with depth and width pruning to meet strict resource budgets. Together with a simplified decoder and an upgraded Dense O2O, this unified design enables DEIMv2 to achieve a superior performance-cost trade-off across diverse scenarios, establishing new state-of-the-art results. Notably, our largest model, DEIMv2-X, achieves 57.8 AP with only 50.3 million parameters, surpassing prior X-scale models that require over 60 million parameters for just 56.5 AP. On the compact side, DEIMv2-S is the first sub-10 million model (9.71 million) to exceed the 50 AP milestone on COCO, reaching 50.9 AP. Even the ultra-lightweight DEIMv2-Pico, with just 1.5 million parameters, delivers 38.5 AP, matching YOLOv10-Nano (2.3 million) with around 50 percent fewer parameters. Our code and pre-trained models are available at https://github.com/Intellindust-AI-Lab/DEIMv2

Paper Structure

This paper contains 13 sections, 1 equation, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Overall architecture.
  4. ViT-based variants.
  5. HGNetv2-based variants.
  6. Spatial Tuning Adapter.
  7. Efficient Decoder.
  8. Enhanced Dense O2O.
  9. Training setting and loss.
  10. Experiments
  11. Comparison to state-of-the-art real-time object detectors.
  12. Comparison to competitive ultra-light object detectors.
  13. Conclusion

Figures (2)

  • Figure 0: Compared with state-of-the-art real-time object detectors on COCO lin2014microsoft, all variants of our proposed DEIMv2 (S, M, L, X) achieve superior performance in terms of average precision (AP) while maintaining similar parameters and less computational cost.
  • Figure 1: Backbone design of our ViT-based variants. We integrate DINOv3 with the proposed Spatial Tuning Adapter (STA).