README.md

---
viewer: false
tags: [uv-script, computer-vision, object-detection, sam3, image-processing]
license: apache-2.0
---

# SAM3 Object Detection

Detect objects in images using Meta's [sam3](https://huggingface.co/facebook/sam3) (Segment Anything Model 3) with text prompts. Process HuggingFace datasets with zero-shot object detection using natural language descriptions.

## Quick Start

**Requires GPU.** Use HuggingFace Jobs for cloud execution:

```bash
hf jobs uv run --flavor a100-large \
    -s HF_TOKEN=HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/sam3/raw/main/detect-objects.py \
    input-dataset \
    output-dataset \
    --class-name photograph
```

## Example Output

Here's an example of detected objects (photographs in historical newspapers) with bounding boxes and confidence scores:

<div style="max-width: 400px;">
<img src="./example-detection.png" alt="Example Detection" style="width: 100%; height: auto;"/>

_Photograph detected in a historical newspaper with bounding box and confidence score. Generated from [davanstrien/newspapers-image-predictions](https://huggingface.co/datasets/davanstrien/newspapers-image-predictions)._

</div>

## Local Execution

If you have a CUDA GPU locally:

```bash
uv run detect-objects.py INPUT OUTPUT --class-name CLASSNAME
```

## Arguments

**Required:**

- `input_dataset` - Input HF dataset ID
- `output_dataset` - Output HF dataset ID
- `--class-name` - Object class to detect (e.g., `"photograph"`, `"animal"`, `"table"`)

**Common options:**

- `--confidence-threshold FLOAT` - Min confidence (default: 0.5)
- `--batch-size INT` - Batch size (default: 4)
- `--max-samples INT` - Limit samples for testing
- `--image-column STR` - Image column name (default: "image")
- `--private` - Make output private

<details>
<summary>All options</summary>

```
--mask-threshold FLOAT       Mask generation threshold (default: 0.5)
--split STR                  Dataset split (default: "train")
--shuffle                    Shuffle before processing
--model STR                  Model ID (default: "facebook/sam3")
--dtype STR                  Precision: float32|float16|bfloat16
--hf-token STR               HF token (or use HF_TOKEN env var)
```

</details>

## HuggingFace Jobs Examples

### Historical Newspapers

Detect photographs in historical newspaper scans:

```bash
hf jobs uv run --flavor a100-large \
    -s HF_TOKEN=HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/sam3/raw/main/detect-objects.py \
    davanstrien/newspapers-with-images-after-photography \
    my-username/newspapers-detected \
    --class-name photograph \
    --confidence-threshold 0.6 \
    --batch-size 8
```

### Document Tables

Extract tables from document scans:

```bash
hf jobs uv run --flavor a100-large \
    -s HF_TOKEN=HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/sam3/raw/main/detect-objects.py \
    my-documents \
    documents-with-tables \
    --class-name table
```

### Wildlife Camera Traps

Detect animals in camera trap images:

```bash
hf jobs uv run --flavor a100-large \
    -s HF_TOKEN=HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/sam3/raw/main/detect-objects.py \
    wildlife-images \
    wildlife-detections \
    --class-name animal \
    --confidence-threshold 0.5
```

### Quick Testing

Test on a small subset before full run:

```bash
hf jobs uv run --flavor a100-large \
    -s HF_TOKEN=HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/sam3/raw/main/detect-objects.py \
    large-dataset \
    test-output \
    --class-name object \
    --max-samples 20
```

### Using Different GPU Flavors

```bash
# L4 (cost-effective)
--flavor l4x1

# A100 (fastest)
--flavor a100
```

See [HF Jobs pricing](https://huggingface.co/pricing#spaces-compute).

## Output Format

Adds `objects` column with ClassLabel-based detections:

```python
{
    "objects": [
        {
            "bbox": [x, y, width, height],
            "category": 0,  # Always 0 for single class
            "score": 0.87
        }
    ]
}
```

Load and use:

```python
from datasets import load_dataset

ds = load_dataset("username/output", split="train")

# ClassLabel feature preserves your class name
class_name = ds.features["objects"].feature["category"].names[0]
print(f"Detected class: {class_name}")

for sample in ds:
    for obj in sample["objects"]:
        print(f"{class_name}: {obj['score']:.2f} at {obj['bbox']}")
```

## Detecting Multiple Object Types

To detect multiple object types, run the script multiple times with different `--class-name` values:

```bash
# Detect photographs
hf jobs uv run ... --class-name photograph

# Detect illustrations
hf jobs uv run ... --class-name illustration

# Merge results as needed
```

## Performance

| GPU | Batch Size | ~Images/sec |
| --- | ---------- | ----------- |
| L4  | 4-8        | 2-4         |
| A10 | 8-16       | 4-6         |

_Varies by image size and detection complexity_

## Common Use Cases

- **Documents:** `--class-name table` or `--class-name figure`
- **Newspapers:** `--class-name photograph` or `--class-name illustration`
- **Wildlife:** `--class-name animal` or `--class-name bird`
- **Products:** `--class-name product` or `--class-name label`

## Troubleshooting

- **No CUDA:** Use HF Jobs (see examples above)
- **OOM errors:** Reduce `--batch-size`
- **Few detections:** Lower `--confidence-threshold` or try different class descriptions
- **Wrong column:** Use `--image-column your_column_name`

## About SAM3

[SAM3](https://huggingface.co/facebook/sam3) is Meta's zero-shot vision model. Describe any object in natural language and it will detect it—no training required.

**Note:** This script uses transformers from git (SAM3 not yet in stable release).

## See Also

More UV scripts at [huggingface.co/uv-scripts](https://huggingface.co/uv-scripts):

- **dataset-creation** - Create HF datasets from files
- **vllm** - Fast LLM inference
- **ocr** - Document OCR

## License

Apache 2.0