Computer vision analysis inputs

The following inputs are required to create a computer vision analysis (either an Object detection or Image classification one): a Managed folders containing the images you want to learn from and a dataset where each row corresponds to an image of your folder, with two columns:

  • A target column containing, for each image, annotations or a category (depending on whether you are doing an Object Detection or Image classification analysis).

  • An image path column specifying the image paths from your folder’s root.

This dataset can be created from the managed folder using the List Folder Contents recipe.

Target column format for Object detection

The target column for Object detection must have, for each image, a json of the following format:

[{
    "bbox": [xmin_bbox, ymin_bbox, width_bbox, height_bbox]
    "category": "cat"
 }, {
    "bbox": [xmin_bbox, ymin_bbox, width_bbox, height_bbox]
    "category": "dog"
}]

The top left of the image having coordinates (0, 0).

If you downloaded your dataset in Pascal or VOC format you can use the plugin Image annotations to Dataset to reformat your annotations and create a Dataset with the right format for computer vision.

If you annotated your images using the plugin ML Assisted Labeling, use the “Reformat image annotations” recipe to create a dataset for Computer vision.

Rows having a bad format or without annotations will be dropped during training.

Target column format for Image classification

The target column for Image classification must be a string or integer for each image (representing its category):

"category 1"

If you have a managed-folder with images being organized in different sub-folders named according to the categories of your dataset, note that you can use the built-in “List Contents” recipe to create the input dataset.

Rows having a bad format or empty rows will be dropped during training.

Image path column

This column contains the relative path of each image from the folder root. Corrupted or missing images are ignored during training.

Supported images formats

Supported image formats for computer vision in DSS are: jpg, jpeg, and png.

There is no image size requirement for using computer vision in DSS. However they should be able to fit into memory (see batch size in Model architectures & training parameters).