Appendix: Training Format

Introduction

The VisionAI format is known for its comprehensive structure, accommodating a wide range of data types and annotations for AI model training. However, to facilitate a more user-friendly experience, we have developed a simplified version of this format for the training process.

Key Features

The simplified training format retains critical components such as image paths, annotations, and class labels, while omitting optional fields that are not directly involved in the training process.

• Streamlined Structure: Focuses on essential elements required for training, reducing complexity.

• Ease of Use: Designed for quick setup and integration with training pipelines.

• Compatibility: Maintains compatibility with standard VisionAI format while simplifying the process.

---

Folder structure

├──dataslice-1
│   ├── images
│   │   ├── img1.jpg
│   │   ├── img2.jpg
│   │   └── img3.jpg
│   └── annotations.jsonl 

---

Contents for annotations.jsonl

Image Classification

Name	Type	Definition	Required
filename	string	The image file name under images folder	true
width	int	Image width	true
height	int	Image height	true
label	int	The index for the label category	true

Example

{"filename": "img1.jpg", "width": 1280, "height": 720, "label": 0}
{"filename": "img2.jpg", "width": 1280, "height": 720, "label": 1}

Box2D object detection

Name	Type	Definition	Required
filename	string	The image file name under images folder	true
width	int	Image width	true
height	int	Image height	true
objects	dict	Contains the detection objects with the key below *label: list of the assigned object category-index *bbox: list of object info which contain x-center, y-center, bbox-width, bbox-height	true

Example

{"filename": "img1.jpg", "width": 1280, "height": 720, 
  "objects": {
    "label": [0, 1, 0, 3],
    "bbox": [[xc, yc, w, h], [xc, yc, w, h], [xc, yc, w, h], [xc, yc, w, h]]
}
{"filename": "img2.jpg", "width": 1280, "height": 720, 
  "objects": {
    "label": [0, 2],
    "bbox": [[xc, yc, w, h], [xc, yc, w, h]]
}