Intellindust-AI-Lab/DEIMv2

[DEIMv2] Real Time Object Detection Meets DINOv3

What it solves

DEIMv2 addresses the need for high-performance, real-time object detection that can scale across different hardware constraints, from ultra-light edge devices to more powerful systems. It aims to provide state-of-the-art (SOTA) accuracy on benchmarks like COCO while maintaining low latency.

How it works

DEIMv2 is an evolution of the DEIM framework that integrates features from DINOv3. It offers a variety of model sizes (Atto, Femto, Pico, N, S, M, L, and X) to balance speed and accuracy. The architecture utilizes different backbones, including HGNetv2 for the lightest versions and DINOv3-based ViTs for the larger, more accurate variants. It employs a HybridEncoder, DEIMTransformer decoder, and a PostProcessor to perform detection.

Who it’s for

This project is designed for AI researchers and developers building real-time vision applications, particularly those targeting edge deployment where a balance between parameter count and inference speed is critical.

Highlights

  • Scalable Model Zoo: Provides a wide range of model sizes from an ultra-light "Atto" version (0.5M parameters) to a high-performance "X" version (50.3M parameters).
  • SOTA Performance: The S-sized model achieves over 50 AP on the COCO benchmark.
  • Edge-Ready: Supports export to ONNX and TensorRT for optimized deployment on edge hardware.
  • Broad Integration: Integrated into Intel® Geti™ and X-AnyLabeling for easier fine-tuning and labeling.

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

    A computer vision framework for pretraining, fine-tuning, and distilling SOTA models like DINOv3 and LTDETRv2 for tasks such as object detection and segmentation.

  • Project
    roboflow/rf-detr

    RF‑DETR is a real‑time transformer‑based vision model suite (detection, segmentation, keypoint) built on a DINOv2 backbone. It ships as the `rfdetr` Python package with several size variants (Nano‑2XL), provides Apache‑2.0 (core) and PML 1.0 (XL/2XL) licenses, and includes benchmark tables showing state‑of‑the‑art accuracy‑latency trade‑offs on COCO and RF100‑VL. Installation is via `pip install rfdetr`; usage is a single‑line `model.predict(...)` followed by optional visualisation with the `supervision` library. The project also offers a NAS pipeline on the Roboflow platform for custom architecture search.

  • Project
    yzfzzz/depth-detect

    A high-performance C++ and TensorRT framework that fuses YOLO object detection with monocular/binocular depth estimation to track objects and estimate their motion states.

  • Project
    RizwanMunawar/yolov7-object-tracking

    A computer vision project that integrates YOLOv7 and YOLOv8 for real-time object detection and tracking across video streams from various sources.