Intellindust-AI-Lab/DEIMv2

[DEIMv2] Real Time Object Detection Meets DINOv3

What it solves

DEIMv2 tackles the need for real-time object detection that is both highly accurate and efficient across a wide range of hardware. Traditional detectors often trade off speed for accuracy, but DEIMv2 provides a spectrum of model sizes—from ultra-light (Atto) to extra-large (X)—so you can pick the right balance for your deployment scenario. It achieves state-of-the-art results on the challenging COCO benchmark, with the S-sized model already surpassing 50 AP, proving that accuracy does not have to be sacrificed for speed.

How it works

DEIMv2 is an evolution of the DEIM detection framework. It leverages rich features from DINOv3, a modern vision backbone, to boost performance. For the smaller variants (S and M), it uses custom ViT-Tiny backbones that are distilled from DINOv3, enabling efficient yet powerful models. The codebase provides configurable training, testing, and fine-tuning pipelines, and models are released on Hugging Face for easy loading. It also includes tools for exporting to ONNX and deploying with TensorRT or OpenVINO, plus integration with Intel Geti for code-free fine-tuning.

Who it’s for

DEIMv2 is aimed at developers and researchers who need accurate object detection in real-time applications—whether on edge devices (using Atto, Femto, or Pico) or on high-end servers (using L or X). It is also suitable for teams that want to fine-tune on custom datasets in COCO format, or that require deployment on Intel hardware via Geti. The project welcomes contributions and provides a straightforward path from pretrained models to custom training.

Highlights

  • State-of-the-art performance on COCO: the S model exceeds 50 AP, and the X model reaches 57.8 AP.
  • Eight model variants (Atto, Femto, Pico, N, S, M, L, X) cover a wide range of computational budgets.
  • Uses DINOv3 features and custom distilled ViT backbones to balance accuracy and speed.
  • Models and code are available on Hugging Face and GitHub, with detailed configuration files.
  • Deployment support: ONNX export, TensorRT (for FP16), and OpenVINO via Intel Geti integration.
  • Integrated with X-AnyLabeling and LightlyTrain, expanding practical usability.

Related

  • Project
    Sompote/DINOV3-YOLOV12

    A hybrid object detection framework combining YOLOv12 and DINOv3 to provide superior accuracy and faster convergence, especially on small or complex datasets.

  • Project
    lightly-ai/lightly-train

    LightlyTrain is a framework for the entire computer vision model lifecycle, enabling users to pretrain, fine-tune, and distill state-of-the-art models for detection and segmentation tasks with support for edge deployment.

  • 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
    Peterande/D-FINE

    D‑FINE is a real‑time object‑detection framework that redefines DETR’s box regression as a fine‑grained distribution refinement and adds a self‑distillation module. It offers several model sizes (‑N to ‑X) that achieve state‑of‑the‑art COCO/AP scores (42.8 %–55.8 %) while running at 70‑470 FPS on a T4 GPU, with no extra inference cost. The repo provides pretrained checkpoints, config files, and full training/evaluation scripts for COCO, Objects365, and custom COCO‑style datasets.

  • Project
    sunsmarterjie/yolov12

    YOLOv12 is an attention-centric computer vision framework that delivers high-accuracy, real-time object detection, segmentation, and classification by optimizing attention mechanisms for speed.