README_custom.md

Custom Dataset

How to Implement Your Own Datset Converter

Create Your Dataset Class

1. Inherit BaseDataFormat class
2. Implement some functions.

• __init__()
  Call _parse_annotation() and _parse_class_list()
  Create MSCOCODataset() instance.
• _parse_annotation()
  Implement annotation parser and set self.image_id_to_image_info and self.image_id_to_annotation_list.

self.image_id_to_image_info[IMAGE_ID] = {
  "file_name": IMAGE_FILE_NAME,
  "file_path": IMAGE_FILE_PATH,
  "width": IMAGE_WIDTH,
  "height": IMAGE_HEIGHT,
}

self.image_id_to_annotation_list[IMAGE_ID].append({
    "annotation_id": ANNOTATION_ID,
    "class_id": CLASS_ID,
    "bbox": BBOX, # Absolute points [left, top, width, height]
    "score": SCORE or None, # If exists, otherwise None
})

• _parse_class_list()
  Implement class list parser and set self.class_id_to_class_name and self.class_name_to_class_id.
• convert()
  Refer to YOLO class
• validation_check()
  Refer to YOLO class

3. Add Supporting Dataset Format Add your own dataset format name to the programs.

• supported_data_format_list
• create_src_dataset_class()