---
title: SAM3 Detection Browser
emoji: 🔍
colorFrom: blue
colorTo: green
sdk: static
pinned: false
models:
- facebook/sam3
---

# 🔍 SAM3 Object Detection Browser

A simple, interactive web application for browsing and exploring object detection results from Meta's [SAM3 (Segment Anything Model 3)](https://huggingface.co/facebook/sam3).

## Features

- **Visual Detection Display**: View images with bounding boxes overlaid on detected objects
- **Confidence Filtering**: Adjust confidence threshold with a slider to filter detections
- **Statistics Dashboard**: See overall detection statistics including:
  - Total filtered images
  - Images with detections
  - Total detections
  - Average detections per image
- **Flexible Navigation**: Browse through datasets with pagination
- **Real-time Updates**: Filters update visualization instantly
- **Color-coded Confidence**: Bounding boxes and scores are color-coded:
  - 🟢 Green: High confidence (≥70%)
  - 🟠 Orange: Medium confidence (40-70%)
  - 🔴 Red: Low confidence (<40%)

## How to Use

1. **Enter a Dataset ID**: Input any HuggingFace dataset that contains object detection results from SAM3
   - Default: `davanstrien/newspapers-image-predictions`
   - Format: `username/dataset-name`

2. **Select Split**: Choose which dataset split to view (train/validation/test)

3. **Adjust Filters**:
   - Move the confidence slider to filter detections by minimum confidence score
   - Check "Show only images with detections" to hide images without any detected objects

4. **Browse**: Navigate through pages using Previous/Next buttons

## Dataset Requirements

This browser works with any HuggingFace dataset that has:
- An `image` column containing images
- An `objects` column with the structure:
  ```python
  {
      "bbox": [[x, y, width, height], ...],    # Bounding box coordinates
      "category": [0, 0, ...],                  # Category indices
      "score": [0.8, 0.6, ...]                  # Confidence scores
  }
  ```

This matches the output format from the [detect-objects.py](https://huggingface.co/datasets/uv-scripts/sam3/blob/main/detect-objects.py) script.

## Example Datasets

- [davanstrien/newspapers-image-predictions](https://huggingface.co/datasets/davanstrien/newspapers-image-predictions) - Photograph detections in historical newspapers

## Creating Your Own Detection Dataset

Use the SAM3 detection script to create your own datasets:

```bash
# Detect objects in your images
uv run https://huggingface.co/datasets/uv-scripts/sam3/raw/main/detect-objects.py \
    your-dataset \
    your-output-dataset \
    --class-name "photograph" \
    --confidence-threshold 0.5
```

Then view the results by entering `your-output-dataset` in this browser!

## Technical Details

- **Technology**: Pure HTML/CSS/JavaScript (no backend required)
- **Data Source**: HuggingFace Datasets API (parquet endpoint)
- **Rendering**: HTML5 Canvas for bounding box overlays
- **Performance**: Loads 12 images per page for optimal performance

## Related Projects

- [SAM3 Detection Script](https://huggingface.co/datasets/uv-scripts/sam3) - Create detection datasets
- [SAM3 Model](https://huggingface.co/facebook/sam3) - The base model
- [UV Scripts Organization](https://huggingface.co/uv-scripts) - More ready-to-run ML scripts

## License

MIT License - Feel free to use and modify for your own projects!