Because active learning involves an iterative annotation of the most-informative images, we assume that the training, validation and test images have their corresponding json/xml annotations from the either the LabelMe software, V7-Darwin software, Supervisely software or CVAT software (when using CVAT, export the annotations to LabelMe 3.0 format). Our default annotation procedure is based on the LabelMe software. Please find the installation instructions of labelme on: https://github.com/wkentaro/labelme.
maskAL will select the most-informative images for you to annotate.
Annotation procedure (LabelMe):
- Annotate each individual object by a polygon. Use the button Create Polygons
- Assign the correct class name to the object.
- When an object consists of two or more separated parts: draw each separated part by an individual polygon and link the polygons by the same group id.
The broccoli head of the class "cateye" is occluded by a leaf, causing two separated instances of the broccoli head. Annotate the individual instances by a separate polygon and link them with an unique group_id (in this example 1). Suppose there is another occluded broccoli head in the image: then use another group_id (for example 2).
