Spaces:
Running on Zero
Running on Zero
| title: DEIMv2-S Object Detection | |
| emoji: 🎯 | |
| colorFrom: gray | |
| colorTo: purple | |
| sdk: gradio | |
| sdk_version: 6.31.0 | |
| python_version: "3.12" | |
| app_file: app.py | |
| startup_duration_timeout: 30m | |
| short_description: DEIMv2-S real-time COCO object detection (DINOv3) | |
| license: other | |
| license_name: deimv2-research-only | |
| license_link: https://github.com/Intellindust-AI-Lab/DEIMv2/blob/main/LICENSE.md | |
| # DEIMv2-S · Real-Time Object Detection Meets DINOv3 | |
| Interactive demo for **DEIMv2-S** — the flagship checkpoint of | |
| [Real-Time Object Detection Meets DINOv3](https://arxiv.org/abs/2509.20787) | |
| ([code](https://github.com/Intellindust-AI-Lab/DEIMv2)). DEIMv2 couples a | |
| DINOv3-pretrained ViT backbone with a **Spatial Tuning Adapter (STA)** and a | |
| simplified DETR-style decoder under **Dense One-to-One matching**; the S model | |
| reaches **50.9 COCO AP at 9.7M parameters**, running end-to-end (NMS-free). | |
| Weights: [Intellindust/DEIMv2_DINOv3_S_COCO](https://huggingface.co/Intellindust/DEIMv2_DINOv3_S_COCO) | |
| (loaded from this Space, no gated access required — the `vit_tiny` backbone is | |
| trained from scratch and fully contained in the checkpoint). | |
| ## How it works | |
| - Preprocessing follows the authors' `tools/inference/torch_inf.py`: | |
| `Resize(640×640)` → `ToTensor` → **ImageNet normalization** (used by all | |
| DINOv3/ViT-backbone variants). | |
| - The full `engine/` package from the official repo is bundled for the model | |
| definition; `app.py` replicates the `DEIMv2` wrapper class from the authors' | |
| `hf_models.ipynb` and loads weights via `PyTorchModelHubMixin`. | |
| - Boxes are scaled back to the **original image size** and scores are sigmoid | |
| confidences; each COCO category is drawn in its own color (per-category HSV | |
| palette from the authors' notebook). | |
| - Runs on ZeroGPU (`spaces.GPU`); inference is a single forward pass of a 9.7M | |
| parameter detector, so latency is well under a second on an A100. | |
| ## Example images | |
| - `examples/authors_example.jpg` — the demo image shipped in the official | |
| DEIMv2 repository (non-commercial research license). | |
| - All other examples are CC0 images from | |
| [linoyts/repo-to-space-example-inputs](https://huggingface.co/datasets/linoyts/repo-to-space-example-inputs). | |
| ## License | |
| DEIMv2 code and weights are released by Intellindust AI Lab under a | |
| **non-commercial research license** — see | |
| [LICENSE.md](https://github.com/Intellindust-AI-Lab/DEIMv2/blob/main/LICENSE.md). | |
| Commercial use requires a separate license from Intellindust. | |
| ## Citation | |
| ```bibtex | |
| @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 preprint arXiv:2509.20787}, | |
| year = {2025} | |
| } | |
| ``` | |