Read dataset object-wise

[MalteEbner]

Description

For object detection tasks, it should be possible to read a dataset object-wise, do the embedding on the objects cropped from the full image and then upload the object to the webapp.

Proposed workflow:

• add cfg parameter path_to_bounding_boxes
• take the images and bounding box information (in yolo-format) and extract the cropped images into a temporary directory
• use this temporary directory as input directory for all following steps: Training, Embedding and Uploading are done object-wise automatically

Open questions

• What is the exact format of the bounding box information? Yolo, but is it a csv of image_name/X1/Y1/X2/Y2 or yaml? What do we support?
• How are the two workflow steps implemented?
  • Option A: The extraction and usage happens under the hood when the cfg param is passed.
  • Option B: There is a separate lightly-crop command needing an output directory, used as input_dir for the following steps.
    This also allows for options like having a cropping padding.

[IgorSusmelj]

Something like this?

We expect the dataset to be in the Yolo format:

`- data_root
  `- images
    `- img_001.jpg
    `- img_002.jpg
  `- labels
    `- img_001.txt
    `- img_002.txt

Then a user can run lightly-crop:

lightly-crop images_dir=data_root/images labels_dir=data_root/labels output_dir=data_root/images_cropped padding=0.1

That would translate into

• get images and labels from the folders
• store cropped images
• add a pudding of 0.1 = 10% to the crops

[MalteEbner]

Ideally use the dataset for MilkyWay for it: https://github.com/eg4000/SKU110K_CVPR19