Is there any way that can show the performance of Cascade Object Detector ?

Hi everyone,

I'm using trainCascadeObjectDetector function to detect multiple objects. The detector can detect almost all of the object, but there are many false positives. I'm trying to find out the reason. Is there any threshold or probability that detector uses to determine whether a region is an object or background? If yes, where can I found it?

Thanks.

Answer 1

[Anand]

I presume you're using vision.CascadeObjectDetector and not trainCascadeObjectDetector for detection. The trainCascadeObjectDetector trains an object detector using training examples you provide and returns an XML file which you can use as input to the vision.CascadeObjectDetector function.

There are a few levers you can play with to tune detection performance:

1. Update 'MinSize' and 'MaxSize' parameters to constrain size of detected objects.
2. Increase 'MergeThreshold', which controls the threshold used to merge multiple near by detections to a single final detection.

You can read about these in the help for vision.CascadeObjectDetector.