Real-time object detection: DEIMv2 meets DINOv3

Abstract: eight models for different devices

Dense O2O gave DEIM faster convergence and improved detection performance. The authors extend this recipe with DINOv3 features to create DEIMv2, spanning eight model sizes from X to Atto for GPU, edge and mobile deployment.

X, L, M and S use DINOv3-pretrained or distilled backbones. The Spatial Tuning Adapter, or STA, transforms DINOv3’s single-scale output into multi-scale features and complements strong semantic representations with fine-grained detection details.

Ultra-lightweight Nano, Pico, Femto and Atto use HGNetv2 with depth and width pruning to meet strict resource limits. A simplified decoder and upgraded Dense O2O combine these choices into a unified family with different performance–cost trade-offs.

Method: context and detail at several scales

The upper branch of the source diagram extracts DINOv3 features. The lower STA branch supplies local detail through convolutional operations and spatial downsampling. Fusion nodes connect the two representations.

The resulting features have scales of 1/8, 1/16 and 1/32 relative to the input image. The Bi-Fusion detail shows bilinear resizing, channel concatenation and a 1 × 1 convolution. This explains the published architecture; it is not a separate Deepvizor experiment.

Semantic features describe what is in the image, while local representations help locate boundaries and smaller elements. STA brings both types of information into downstream detection.

Results: detection quality and model size

DEIMv2-X reaches 57.8 AP with 50.3 million parameters. DEIMv2-S achieves 50.9 AP with 9.71 million parameters, crossing the reported 50 AP threshold below 10 million parameters.

DEIMv2-Pico uses 1.5 million parameters and reaches 38.5 AP. The source compares this AP with YOLOv10-Nano at 2.3 million parameters. The table preserves the three variants explicitly reported in the abstract.

AP is a detection metric on the COCO benchmark, not the percentage of correctly handled industrial alarms. The chart compares AP with parameter count; it does not report latency on Deepvizor devices.

Original AP-versus-parameters chart, with curves and values preserved.

Open research figure ↗

Deepvizor note: choosing a model size

Start with the target event and camera conditions. A small distant object and a large object near an entrance may require different input resolutions and quality margins. An algorithm name alone does not determine processing load.

Measure the complete pipeline during a pilot: stream ingestion, decoding, inference, event rules and notification. Compare models on identical video and the intended hardware, considering memory, numerical precision and simultaneous camera count alongside parameters.

Original paper and code

Code and pretrained weights are available in the DEIMv2 repository. The source project and full paper are linked above; the authors’ citation follows.

@article{huang2025deimv2,
  title={Real-Time Object Detection Meets DINOv3},
  author={Huang, Shihua and Hou, Yongjie and Liu, Longfei and Yu, Xuanlong and Shen, Xi},
  journal={arXiv},
  year={2025}
}