2

I'm currently working on a project to explore various means of human detection. To formulate a base-line for comparison, I am looking to utilize several standard detection methods (ie, HOG with SVM, Viola-Jones). As I am using the Caltech Pedestrian Dataset(http://www.vision.caltech.edu/Image_Datasets/CaltechPedestrians/) for testing, I am doing all my work in Matlab to take advantage of their additional toolboxes.

I'm currently stuck on the Viola-Jones problem. The detector evaluation tool requires [frame, x, y, h, w, score] to generate ROC curves for the detectors. I had planned to use the CascadeObjectDetector with the OpenCV trained fullbody.xml model with the following code:

detector = vision.CascadeObjectDetector(MODEL)

BBOX = step(detector,I)

However, the detector only gives the Bounding Box values and no score. Is there some work around to generate ROC curves for Cascade Classifiers? Or another implementation that can provide the desired results?

Dima
  • 38,860
  • 14
  • 75
  • 115
ChrisD
  • 23
  • 3

1 Answers1

1

Unfortunately, there is currently no way to get a confidence score from vision.CascadeObjectDetector.

Dima
  • 38,860
  • 14
  • 75
  • 115
  • That's basically what I discovered. It doesn't make sense to me why any detector wouldn't produce some confidence metric, but thats just me. I ended up using OpenCV's Cascade People Detector which recorded the values needed to make a ROC plot. – ChrisD Apr 03 '15 at 19:12
  • I was super excited to try Matlab's Computer vision toolbox but this is a very serious limitation! – user42174 Nov 02 '15 at 20:05
  • side note: I ended up generating bboxes for continuously varying "thresholds" (the MergeThreshold parameter), with both positive and negative exemplars, making sure that the variations in thereshold are sufficient to get the (0,0) and (1,1) points of the ROC curve... – user42174 Nov 02 '15 at 20:39