How to take advantage of Geti's active learning model when labeling many classes?

[charecktowa]

I'm inexperienced with data labeling and managing large datasets.

We're using Geti's active learning approach. However, we've run into a problem: when labeling multiple species (the ones in the images suggested from the active set), the model tends to suggest multiple incorrect labels (which totally makes sense). This means we often have to either manually delete the incorrect suggestions or start over with manual labeling, which defeats the purpose of using AI for labeling.

I'm unsure if this is the best place for this question, but I'm seeking advice on a more effective strategy to leverage Geti's AI-assisted labeling capabilities. We've found Geti to be faster than other platforms, and one of the main reasons we chose it was the AI labeling out of the box.

Currently, we have shifted our approach: we are now labeling all images belonging to a single species at a time. We already know which species is present in each image, but we need to create the bounding boxes for object detection.

[rajeshgangireddy]

Hi @charecktowa
What you're experiencing is a common issue in multi-class object detection when the model isn’t confident or well-trained enough across all classes. In Geti, the top active learning images are the "hard" samples that the model finds confusing between classes and hence needs manual labelling. So, it is expected that some of them are mislabelled.

Currently, we have shifted our approach: we are now labeling all images belonging to a single species at a time. We already know which species is present in each image, but we need to create the bounding boxes for object detection.

This is a good approach. You can select the label (which will remain unchanged as you are labelling the same species) and use the annotation tool that creates automatic bboxes when hovered on an object. This makes annotating faster.

The other approach is to train in batches - annotate a few images such that there are enough objects for each class. Once you label such a batch of images, retrain the model. The model’s predictions should improve as each species is learned well. The AI predictions for auto labelling would also get better when you start annotating after the training is finished.
How many images to train depends on how similar your dataset classes are.

Let me know if you have more questions.