Lightly Object Detection Format

Overview

The Lightly format is designed for efficient handling of object detection predictions in machine learning workflows. It provides a straightforward structure that's easy to parse and generate. For detailed information about the prediction format, refer to the Lightly AI documentation.

Specification of Lightly Detection Format

The format uses a JSON file per image containing: - file_name: Name of the image file - predictions: List of object detections - category_id: Integer ID of the object category - bbox: List of [x, y, width, height] in absolute pixel coordinates - score: Optional confidence score (0-1)

File Structure

dataset/
├── images/
│   ├── image1.jpg
│   └── image2.jpg
└── predictions/
    ├── image1.json
    └── image2.json

Example

{
  "file_name": "image1.jpg",
  "predictions": [
    {
      "category_id": 0,
      "bbox": [100, 200, 50, 30],
      "score": 0.95
    },
    {
      "category_id": 1,
      "bbox": [300, 400, 80, 60],
      "score": 0.87
    }
  ]
}