---
license: apache-2.0
tags:
  - onnx
  - object-detection
  - rf-detr
  - pet-detection
library_name: onnx
pipeline_tag: object-detection
---

# RF-DETR Small (ONNX) — pet detection for Gallery

ONNX export of [RF-DETR](https://github.com/roboflow/rf-detr) Small by Roboflow, used by
[Gallery](https://github.com/open-noodle/gallery) for pet detection.

Unmodified COCO-pretrained weights. No fine-tuning.

## Files

| Path                   | Description                      |
| ---------------------- | -------------------------------- |
| `detection/model.onnx` | RF-DETR Small, opset 17, batch 1 |

## Inference contract

Getting any of this wrong degrades output silently rather than erroring.

**Input** — `input`, shape `[1, 3, 512, 512]`, float32:

1. Decode to **RGB** (not BGR)
2. Resize to 512×512 — plain square resize, **not** letterboxed
3. Scale to `[0, 1]`
4. Normalise with ImageNet statistics: mean `[0.485, 0.456, 0.406]`, std `[0.229, 0.224, 0.225]`
5. Transpose HWC → CHW, add batch dimension

**Output** — two tensors:

| Name     | Shape          | Meaning                                        |
| -------- | -------------- | ---------------------------------------------- |
| `dets`   | `[1, 300, 4]`  | Boxes as cx, cy, w, h — normalised to `[0, 1]` |
| `labels` | `[1, 300, 91]` | Class logits, **pre-sigmoid**                  |

Apply sigmoid to `labels`, then threshold. Classes use the **91-class COCO id space**
(90 categories plus background), not the contiguous 80-class space YOLO uses — so
`bird=16, cat=17, dog=18, horse=19, sheep=20, cow=21`.

The 300 queries are already deduplicated. **No NMS step is required.**

## License

Apache-2.0, inherited from RF-DETR. See the
[upstream repository](https://github.com/roboflow/rf-detr) for full terms.