Splitting_Labeled_Images.md

Splitting Labeled Images

Labeling the original large images and then splitting them afterwards has the benefit, that the splitting can be prformed multiple times to try out different resolutions. It can also give the labeler more context when labeling.

The steps are:

• export in CVAT as COCO 1.0

• split images with the sahi library

  sahi coco slice  --image_dir path/to/images/ --dataset_json_path path/toannotations.json --slice_size 640 --output_dir tiled_coco_dataset/

  This will split the images and also transform the labels for the new images.

• convert to a YOLO dataset with convert_coco

      from ultralytics.data.converter import convert_coco
      convert_coco("/path/to/annotations.json/parent/dir/")

  This will create a .txt file for each image in your dataset. Create a new directory for and move the images and labels into it. eg:

  • yolo_dataset/images: for all the images.
  • yolo_dataset/labels: for all the label .txt files.

• split into train val dataset with autosplit

  from ultralytics.data.utils import autosplit
  autosplit( path="./yolo_dataset", weights=(0.85, 0.15, 0.0), annotated_only=False )

  This will create a autosplit_train.txt and autosplit_val.txt in your dataset folder defining your train/val split.

The final dataset then looks like this:

• Folderstructure

      yolo_dataset/
      |-images/
      |... all images
      |-labels/
      |... all labels .txt, same name as the image
      - autosplit_train.txt
      - autosplit_val.txt
      - data.yaml
  

• data.yaml

  names:
    0: blue box
    1: light pole
  path: /path/to/yolo_dataset
  train: autosplit_train.txt
  val: autosplit_val.txt

• autosplit.txt

  ./images/imagename.jpg
  ...

  The path should be relative to path: /path/to/yolo_dataset and start with ./

Inferencing with a sliding window

When inferencing is done on the large image directly the image is scaled down. This can cause bad results if the model was trained on splitted image, because the inferencing resoluton is now a lot lower than the training resolution.

The sahi library can automatically perform sliding window inferencing on larger images:

• documentation
• example src/predict_tiled.py