How can I incorporate extractHOGFeatures with trainCasca​deObjectDe​tector?

I've managed to used trainCascadeObjectDetector to detect my object. I am planning to incorporate extractHOGFeatures together with trainCascadeObjectDetector to improve my detection accuracy. May I know does this help to improve the detection? If so, how do I do it?

Answer 1

[Dima Lisin]

Hi Leonard,

You cannot use extractHOGFeatures together with trainCascadeObjectDetector. However, you can specify the feature type for trainCascadeObjectDetector to use, and one of the feature types it supports is HOG.

I cannot tell you whether using HOG will improve your accuracy. That mainly depends on the type of objects you are trying to detect. What I can tell you, is that training with HOG features is much faster and takes much less memory.

[Leonard Yeo]

I'm actually trying to detect wheelchairs. I'm currently using haar feature to train with trainCascadeObjectDetector however, the detection rate is about 60-65% and consist of many false detection. I've tried using HOG feature but it did not manage to detect anything as compared to haar which was quite surprising because shouldn't HOG perform better than haar. Are there any other methods to improve my detection rate? I've tried adding more images to it but the detection rate only improves a little.

[Dima Lisin]

HOG is not necessarily better than Haar. For faces, Haar seems to do better, for detecting people - HOG seems to do better. What is better for wheelchairs is anyone's guess. :(

As I said in the other question, try splitting up your images into sub-categories like "front view", "side view", "back view", and try training separate detectors for each one.

By the way, 600 positive samples is actually not that many. You may need more.