Cvat: Active Learning and Online Learning Functionality

[davodogster]

Hi guys,

Are you planning on implementing support for Active Learning or Online Learning for fast object detection labeling in the near future? Cheers, Sam

[snosov1]

Hey, @davodogster !

Currently, we're working on a POC for "regular" learning with CVAT-based annotation. Online/active learning is a next logical step, but, I think, we'll get around it no sooner than 2021.

[davodogster]

Great to hear that. Does the "regular learning" involve training a model inside CVAT and using it to predict unlabeled images?

[snosov1]

Not initially. "Regular learning" means:

1. Setting up a task in CVAT with a predefined annotation schema (e.g. a bounding box annotation with a single label "face" to train a face detector).
2. Exporting the annotated tasks as a dataset
3. Training the model using the exported dataset.

Inclusion of the model into CVAT to annotate unlabeled data (and to fine-tune again and to annotate more and so forth) is what I'd call "online learning" and it's something further down the road.

[davodogster]

@snosov1 That makes sense - thanks for the clarification. Looking forward to the online learning in the future. I see that LabelStudio also have it and active learning on their todo list.