Training a soft cascade classifier

edit

Ah, didn't know about that one. According to the documentation it seems to be pretty fast and the 2nd paper "Pedestrian detection at 100 frames per second." indicates that too, or is it not speeded up?

The 100 Hz detector is developed at the computervision research group I am associated with at our University. To actually get those speeds, you need to combine the softcascade with the stixels approach and use ground plane assumptions. I do not think that this will be completely implemented in OpenCV, only the softcascade part.

The problem is that I need to create a cascade of boosted classifiers to verify my own features, and, as far as I know, OpenCV only provides this application to be used with the Haar-like features using the Viola-Jones approach.

My alogrithm performs nicely in the pedestrian datasets i've tried, but the problem is that I am using a single boosted classifier for decision-making, so I need to evaluate a huge amount of features to get a decision, and this takes about 20 seconds per image. What I need is a number of boosted classifiers that grow in complexity as a sample advances in the structure, so I don't need to evaluate all the features.

Can you help me some how?

What you just described is the links that guanta provided. Just select 20 stages or such as stage parameter, which indicates the amount of weak classifiers.

Hey, @StevenPuttemans! I have one question about this paper. I thought that the Soft Cascade by itself provided a speedup of 20 times (Table 1 of the paper). The impression it gives is that by using only Soft Cascade we already can achieve this 20x speed up. Using the stixels and ground plane assumptions, we can increase this speed up even further. Is that correct? I mean, shouldn't the speed up provided by the Soft Cascade be independent of the other optimization approaches?

Ofcourse, the softcascade can increase a speedup on itself. However the increase reported is on the test application provided. It all depends on the amount of weak features you want to evaluate. It is correct that the softcascade can introduce an optimization without the ground plane and stixel assumption. However to reach the full 100Hz detector speed, you need the three parts combined.

So basically my remark is incorrect. At that time I didn't look at the paper well enough. Have done it for my own research now and it is faster due to faster early rejecting instead of depending on stages with multiple feature evaluation.

Hi Steven, could you provide or recommend some links and directions on how to train and use soft cascade? I didn't find related documents in it. Thanks!