A quick way to distinguish similar objects from each using separate objects on one object

Question

I made a classifier (based on HoG functions) that can recognize large vehicles (buses and trucks). But I also want to distinguish between buses and trucks. This causes problems because both cars are large and long. Here is an example from my training data:

As you can see, one truck and one tire are considered from the same point of view, but the classifier does not consider them different.

Given that I have already built my classifier, is there a simple way (without restoring the existing classifier) ​​that I could add as a second step to distinguish between a truck and a bus?

I was thinking about how the SIFT function combines ... to grab this protruding truck head in front. But I have not used it before and was not sure if it is applicable here.

Answer 1

I understand that your current detectors are trying to distinguish between buses and trucks from other objects. Assuming that it distinguishes other objects well and has problems that differ between tires and trucks, you can add a specialized classifier to it.

The purpose of the second classifier should be different between the buses and the truck, given the first classifier. Therefore, you must train its entities, which consider the truck or bus as the first classifier (ignore buses and the truck that are not identified by the first classifier). Given that the samples use their true classification as a concept (rather than the prediction of the first classifier). By doing so, you force the second classifier to focus on the boundaries between buses and trucks. Within these boundaries, functions such as size become irrelevant, since both tires and trucks are large, and therefore the classifier will be forced to find other relevant functions.

After that, you will make a classifier, and if the first classifier displays either a bus or a truck, you must return the result of the second classifier.

This method is actually a special case of boosting , similar in spirit to the Schapire algorithm from Strong weak learning ability