Dataset Management

The initial step in detecting Your Kind of Waste

Image preprocessing

...

Create Dataset

Create a new Dataset

Import Dataset

Requirements
  • The dataset must be in a .zip file format.
  • Annotation files must be named annotations_coco.json.
Warning

MMDetection COCO Format
same as the standard COCO Format but without the NULL Class.
Official format documentation

You can import datasets in two ways: Split Dataset or Non-Split Dataset.

Split Dataset

For a split dataset, the .zip file must contain the following structure:

  • Three folders named exactly:
  • train
  • valid
  • test
  • Each folder should include:
  • The images for the respective split.
  • An _annotations.coco.json file.
Non-Split Dataset

For a non-split dataset, the .zip file should contain:

  • All images in the root directory.
  • A single _annotations.coco.json file in the root directory.

Manual Image Upload

  • Upload images one by one
  • Through the Edit Dataset page
    • Click on the Add new dataset image
    • Choose an image on your device
    • Choose the part of the image that you want to use
Important

Images must be at least 1280x1280 pixels for now, will become a dataset setting

Prepare and Annotate Procedure Best Practices

  • Through the Edit Dataset page

    • Click on the Edit Image button while hovering over an image
    • Check Enable Image checkbox
    • Press the Edit button

  • Through the Image Annotation page

    • Click the Change Status button while the Status Radio Button is set to Enabled

  • Optionally: Enable all images at once

  • Check the annotations in each image, by going throught each image separately and enabling it if satisfactory.

  • Leave Images with annotations which contain annotations about which you are unsure of as disabled images.

  • Consult team about the annotations in the remaining disabled images

    • think about end goal (what will be detected)

Publish Dataset

Preconditions for Dataset Publishing

  • All Images Must Be Enabled
    • All images associated with the dataset must be enabled. (Optionally: Enable all images at once)
  • All Enabled Images Must Have Annotations (Planned Change)
    • In the current implementation, all enabled images in the dataset must have at least one annotation associated with them.
  • Minimum Number of Images Required
    • The dataset must contain at least 100 images to be eligible for publishing.
  • Minimum Number of Classes Required
    • The dataset must include at least 1 dataset class linked to it.

Delete Image

  • Through the Edit Dataset page
    • Click on the Edit Image button while hovering over an image
    • Press the Delete button
    • Press the Delete button again in the new pop-up

Export Dataset

  • Export Dataset for further use
    • Press the Export Annotations (will be changed) to COCO Format button
    • Choose the options that suits your case
      • Export Options
        • All Images default option
        • Only Enable Images toogle toogle
      • Export as Split Dataset
        • Export as Single Directory default option
          • all images are in a single directory
          • a single annotations_coco.json file for all images
        • Export as Split Directories toggle toggle
          • Split Directories
            • train
            • valid
            • test
          • images are split randomly into the three directories
          • each directory has a separate annotations_coco.json file for the images in that directory
          • current implementation works only for single class datasets
Tip

For more information, check the Dataset Management Documentation here.