Long-time lurker- first time poster... first off let me say what a great forum for new OpenCV developers like myself. It's been invaluable. Keep it up- you may not think people actually use your advice, however I can confirm your comments/examples/best practices have saved me a lot of head-aches. (Gold-stars all around)
QUESTION: Which HOGDescriptor SVM performs best for speed/accuracy detecting pedestrians from a fixed point/mono?
I'm looking to improve accuracy re: pedestrian detection using various methods. (aren't we all) Wondering if anyone has done any substantive comparisons between the various SVM options used as HOGDescriptor? Which is faster or more accurate?
THOUGHTS:
getDaimlerPeopleDetector == prebuilt SVM based on the Daimler Pedestrian Data here. It's pre-built/based on a well known data set and seems to perform well, however it is only valid for 48x96 images.
getPeopleDetector48x96 == prebuilt SVM. Not sure the source of training material. Seems to be very fast with no problems so far - good accuracy, however limited to images of only 48x96.
getPeopleDetector64x128 == prebuilt SVM. Again not sure of the source of the training data- same issue 64x128 images. Doesn't detect smaller pedestrians- i.e. field of vision reaches the horizon quickly.
detectMultiScale == prebuilt SVM. Don't know much about this one at all. Anyone else?
Build your own? I've actually downloaded all the Daimler data and mixed it with my own to build a custom SVM. It works well, however the 'weighting' gets a little tricky- Daimler data has over 10,000+ images - my custom images/data is only around 3000.
I'm happy with the results thus far- just looking to improve accuracy/speed. All reply's appreciated.
JTM