A fast CLI tool for image classification using MegaDetector v1000, designed for culling large photo collections.

• Object Detection: Uses MegaDetector v1000-larch for 3-class detection (animal, person, vehicle)
• RAW Support: Handles CR2, CR3, NEF, ARW, DNG files (extracts embedded JPEG preview)
• Quality Scoring: Composite IQA score (sharpness, exposure, contrast, resolution)
• Flexible Filtering: Filter by class, quality threshold, with inverse support
• JSON Output: Machine-readable output for integration with other tools
• Debug Mode: Saves annotated images with bounding boxes
• Pipe-friendly: Default output is just paths for easy shell integration

• Rust 1.70+ (install via rustup)

git clone <repo-url>
cd bof5
cargo build --release
# install locally
cp target/release/bof /usr/local/bin/

Model is not included : it's downloaded from a repo, and converted. See ./model/README.md

cargo run --release -- -t 50 --json ./public/images

bof [OPTIONS] <PATH>

• <PATH> - Path to image folder

# Basic detection - outputs paths only (pipe-friendly)
bof ./photos

# Verbose output with details
bof -v ./photos

# Filter animals and copy to best/ folder
bof -f animal ./photos | xargs -I {} cp {} ./best/

# Save debug images with ROI boxes
bof --debug ./debug_output -f animal ./photos

# JSON output for scripting
bof --json ./photos > results.json

# High quality animals only (quality score >= 60)
bof -f animal -t 60 -v ./photos

# Exclude persons from results
bof -f person --inverse-filter ./photos

# Get low quality images (below threshold 30)
bof -t 30 --inverse-threshold ./photos

• Standard: JPG, JPEG, PNG, WebP, TIFF
• RAW: CR2, CR3 (Canon), NEF (Nikon), ARW (Sony), DNG

image1.jpg
image2.CR3
subfolder/image3.png

image.jpg:
  animal Q:72 [RES:85 CONF:90 SHARP:65 EXP:80 CONTR:70]

image.jpg

Verbose info goes to stderr, paths to stdout for clean piping.

[
  {
    "path": "image.jpg",
    "detections": [
      {
        "class": "animal",
        "confidence": 0.902,
        "quality": {
          "score": 72.0,
          "sharpness": 65.0,
          "exposure": 80.0,
          "contrast": 70.0,
          "resolution": 85.0
        },
        "bbox": { "x": 1912, "y": 1106, "width": 994, "height": 1536 }
      }
    ]
  }
]

1. Preprocessing: Images resized to 640x640, normalized to [0,1]
2. Inference: MegaDetector v1000-larch ONNX model via ONNX Runtime
3. Post-processing: Non-Maximum Suppression (IoU threshold 0.45)
4. Quality Scoring: Composite score from sharpness, exposure, contrast, and resolution

The model detects 3 classes: animal, person, vehicle.

• yolo-rust-ort - YOLOv8/v10 in Rust
• yolo-rs - YOLO inference crate
• ort - ONNX Runtime Rust bindings

MIT