Open dlangenk opened 3 months ago
MAIA currently has an additional step to review the training proposals of the novelty detection or the existing annotations of knowledge transfer. With zero shot detection we could show a subset of detections as "training proposals" and then implement a new way to restart the MAIA job with new parameters.
We could add a zero shot object detection pipeline, which works quite great for more common objects like fish. We could integrate this into the MAIA tool as well, BUT one problem is that the threshold values are rather sensitive, so seeing the results on one image and then finetuning the threshold would be very beneficial. However, this would lead to more implementation effort. The python scripts are already there to perform the detection.