---
pipeline_tag: object-detection
license: apache-2.0
base_model: PekingU/rtdetr_r50vd_coco_o365
library_name: zeromodels
tags:
- keras
- zeromodels
- rt-detr
- detr
- object-detection
- arxiv:2304.08069
- pytorch
- jax
- tf
---

## ***See [our collection](https://huggingface.co/collections/zeromodels/rt-detr-v1-and-v2-6a8eaf859a918a955e683778) for all versions of RT-DETR.***

# Run RT-DETR with Keras 3: JAX, PyTorch, or TensorFlow

[![GitHub](https://img.shields.io/badge/GitHub-ZeroModels-black?logo=github)](https://github.com/IMvision12/ZeroModels) [![Docs](https://img.shields.io/badge/Docs-RT--DETR-blue)](https://imvision12.github.io/ZeroModels/rt_detr/) [![Collection](https://img.shields.io/badge/HF-RT--DETR%20collection-yellow)](https://huggingface.co/collections/zeromodels/rt-detr-v1-and-v2-6a8eaf859a918a955e683778)

# zeromodels/rtdetr-r50vd-coco-o365

Paper: [DETRs Beat YOLOs on Real-time Object Detection (arXiv:2304.08069)](https://arxiv.org/abs/2304.08069) · [HF Papers](https://huggingface.co/papers/2304.08069)

RT-DETR was the first DETR-style detector to beat YOLO on the real-time speed/accuracy tradeoff. It pairs a ResNet-vd backbone with a hybrid encoder that decouples intra-scale attention from cross-scale fusion, then feeds IoU-aware selected queries into a deformable decoder. It is NMS-free: a fixed set of queries, constant inference cost, no NMS threshold to tune.

For more details on the model, please go to PekingU's original [model card](https://huggingface.co/PekingU/rtdetr_r50vd_coco_o365).

Pure-**Keras 3** conversion of [`PekingU/rtdetr_r50vd_coco_o365`](https://huggingface.co/PekingU/rtdetr_r50vd_coco_o365) for [zeromodels](https://github.com/IMvision12/ZeroModels). One implementation runs unmodified on **TensorFlow / Torch / JAX**.

This is an **object detection** checkpoint (`RTDETRDetect`) on COCO (ResNet-50-vd).

## ✨ Quick start

```python
import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from PIL import Image
from zeromodels.models.rt_detr import RTDETRDetect, RTDETRImageProcessor

model = RTDETRDetect.from_weights("zeromodels/rtdetr-r50vd-coco-o365")
processor = RTDETRImageProcessor.from_weights("zeromodels/rtdetr-r50vd-coco-o365")

image = Image.open("your_image.jpg").convert("RGB")
inputs = processor(image)
output = model(inputs["pixel_values"], training=False)
results = processor.post_process_object_detection(
    output, threshold=0.5, target_sizes=[(image.height, image.width)]
)[0]
for score, name, box in zip(
    results["scores"], results["label_names"], results["boxes"]
):
    print(f"{name}: {float(score):.3f} {box}")
```

Load any RT-DETR v1 variant the same way with `from_weights("zeromodels/<variant>")`:

| Variant | Hub | Backbone |
|---|---|---|
| `rtdetr-r18vd` | [`zeromodels/rtdetr-r18vd`](https://huggingface.co/zeromodels/rtdetr-r18vd) | ResNet-18-vd |
| `rtdetr-r18vd-coco-o365` | [`zeromodels/rtdetr-r18vd-coco-o365`](https://huggingface.co/zeromodels/rtdetr-r18vd-coco-o365) | ResNet-18-vd (COCO+O365) |
| `rtdetr-r34vd` | [`zeromodels/rtdetr-r34vd`](https://huggingface.co/zeromodels/rtdetr-r34vd) | ResNet-34-vd |
| `rtdetr-r50vd` | [`zeromodels/rtdetr-r50vd`](https://huggingface.co/zeromodels/rtdetr-r50vd) | ResNet-50-vd |
| `rtdetr-r50vd-coco-o365` | [`zeromodels/rtdetr-r50vd-coco-o365`](https://huggingface.co/zeromodels/rtdetr-r50vd-coco-o365) | ResNet-50-vd (COCO+O365) |
| `rtdetr-r101vd` | [`zeromodels/rtdetr-r101vd`](https://huggingface.co/zeromodels/rtdetr-r101vd) | ResNet-101-vd |
| `rtdetr-r101vd-coco-o365` | [`zeromodels/rtdetr-r101vd-coco-o365`](https://huggingface.co/zeromodels/rtdetr-r101vd-coco-o365) | ResNet-101-vd (COCO+O365) |

## Tips

- Set `KERAS_BACKEND` **before** importing Keras / zeromodels.
- `RTDETRImageProcessor` keeps `do_normalize=False` by default (rescaled `[0, 1]` input, matching upstream).
- See [RT-DETR docs](https://imvision12.github.io/ZeroModels/rt_detr/) and [Loading Weights](https://imvision12.github.io/ZeroModels/loading_weights/).
- Community / upstream safetensors still work via the `hf:` prefix, e.g. `RTDETRDetect.from_weights("hf:PekingU/rtdetr_r50vd_coco_o365")`.

## Special Thanks

A huge thank you to the RT-DETR authors (Baidu / PekingU) for creating and releasing these models.

License: Apache 2.0.