about the function trainCasca​deObjectDe​tector

It seems that I can't feed hand annotated negative samples to the engine as that only picture scenes without the object to be detected can be feed to the engine then the negative samples will be automatically generated.

However, consider the situation where I want to distinguish stop sign from other traffic signs while they all have similar circular contours. If I merely use automatically generated negative samples, there will be a lot of false positives when I only want to detect the stop sign. I figure the positive samples are ample (over 80 positive samples with hand annotation), it's got be something wrong with the negative sample. If I can use hand annotated negative samples, things might change. Is the true? How can I accomplish that?

Answer 1

[Dima Lisin]

Hi Sun,

You are correct. trainCascadeObjectDetector generates negative samples automatically from the negative images. The reason for doing that is precisely to avoid the false positives. All you have to do is give the function images that do not contain the stop signs, but that do contain other signs. trainCascadeObjectDetector will learn to not classify those as stop signs on its own.

Also, 80 positive samples is not very many. You may want to consider labeling more positive samples, or generating more positive samples by applying small transformations and/or adding noise to the ones you already have.