I want to detect bananas using vision.cascade object detector. Is that possible?
Best Answer
Yes, it is possible. You would have to train your own banana detector model using thetrainCascadeObjectDetector function. This will produce an XML file, which you can then use withvision.CascadeObjectDetector. See theTrain a Cascade Object Detector tutorial.
There are multiple things you can try to fix this problem:
The simplest approach is to enlarge the box returned byvision.CascadeObjectDetector by some fixed amount. If you are lucky, you may find the number that makes the box big enough most of the time.
Since you are working with a video, you can track a detected mouth, rather than re-detect it in every frame. Seethis example. You may have to modify the code not to compute a transformation between tracked points, because your deformation is non-rigid.
A much more labor intensive approach would be to train your own mouth detector usingtrainCascadeObjectDetector that is tailored specifically to your data. You can use thetrainingImageLabeler app to create a set of training instances.
Unfortunatelyvision.CascadeObjectDetector does not return a confidence score, and there is no workaround. The reasonvision.PeopleDetector does return a score, is because it is using a SVM classifier, which provides a score.vision.CascadeObjectDetector, on the other hand, uses a cascade of boosted decision trees, which does not lend itself well to computing a confidence score.
Best Answer