I create a Polimago Search Project. When I use Search On New Image to test my classifier. I found some wrong place to find. Can I mark the wrong place not to target
, that make my classifier more accurate.
Since the correct found search results and your false positives look so similar, I don’t see how you can eliminate those false positives just by tuning the search classifier.
What you could do is, leave the search as it is and create a training set with all your search results. Then you split them in two classes: the search results you like and the ones you want to eliminate. If these are distinct enough, you will be able to create a robust classification predictor.
Then you can first apply search and then classify all potential search results to eliminate the false positives. You might be able to eliminate a few but most likely not all of your false positives since they look quite similar at times.
3 Likes
Adding to Franks answer, your results at the corne with holes/bumps seem to be the same except rotation. You can eliminate these by setting rotation invariant off.
Your Example looks like you want to find scratches or faulty spots, is that right? Like Frank said, from the picture you provided it is hard to tell differences and what you are looking for, if you provide some more information it will be easier to assist.
P.s: You can change the invariant type as shown in the picture. Underneath you can also define rotation invariance.
4 Likes