Using custom trained Yolov3 weights to label new images

Question

I've recently trained a custom yolov3 model to recognize animals. The first dataset only had around 900 photos. I have access to 10Ks of photos, but labeling them all through labelImg is going to take ages.

Is there a way to use the pre-trained weights to auto label new photos, with an option to then edit the boxes and tags if it is wrong?

Answer 1

You can maybe predict on the other images using your trained model, see this issue.

Afterwards, these detections can be loaded into labelImg for editing.
Since darknet saves detections into a txt file with just bounding box info, you have to first convert these into Pascal VOC format.
You can do this yourself of course, but people have built their own implementations already (not tested).

Answer 2

If you have a budget there are paid tools like Segments AI and V7's Darwin , both of which appear to have AI that assists with annotation. They learn from the items you have already annotated to automate future annotations. For the record I am affiliated with neither, and I have not used them.