---
title: Anything Detector
emoji: 🔍
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: "5.49.1"
python_version: "3.10"
app_file: app.py
pinned: false
startup_duration_timeout: 1h
---

# 🔍 Anything Detector

Custom object detection and counting without YOLO.

## Workflow

1. Define classes in `training/classes.txt`.
2. Upload training images.
3. Select an image.
4. Draw bounding boxes directly on the image with the mouse.
5. Choose a class for each box.
6. Save annotations.
7. Train RT-DETR.
8. Upload new images and detect/count objects.

The annotation tool is implemented directly with HTML5 Canvas inside Gradio. It does not use the `gradio_image_annotation` package, avoiding the Node.js frontend dependency that can cause Space startup problems.

## Example classes

```text
cornetto
magnum
correto
cone
cup
```

Or anything else:

```text
car
truck
person
dog
cat
```

## Important

For production, attach persistent storage to the Space so the dataset and trained model survive rebuilds.

## Startup fix
The annotation JavaScript is registered at the Gradio Blocks level rather than using the unsupported `js_on_load` argument on `gr.HTML`.


## GPU
This Space is configured for Hugging Face ZeroGPU. GPU-dependent training and inference functions use `@spaces.GPU`; training runs synchronously inside the GPU allocation so a background thread cannot outlive the allocation.