I've a large set of image data completely unlabeled(around 500 images,each containing around 80-100 objects to be labeled). I'm using an annotation tool to label objects in those images for preparing ground truth so that I can check the efficiency of my ML alorithm after applying to those images. Is there any way I can automate this Ground truth preparation so that I dont have to label 50,000 objects manually (by drawing polygons around each of those objects).
Idea/Tool that might work for me:
- A semi-supervised learning method - I will first label few objects, then apply the semi supervised learning algorithm on the unlabeled images. This will obviously do misclassifications. So I will rectify label on wrongly labeled data manually.
- An AI-powered Labeling tool - A already existing tool to intelligently label data, which will do the process as described in above step. ;-)
My primary doubt is how can I rectify the mislabeled data generated after testing and use those for next iteration of training.