Dlib's simple_object_detector() is running slowly

Question

I have trained a SVM model for simple_object_detector(). But while inference on a video it is becoming too slow.

I went through a similar question: Why is dlib so slow finding an object? where an answer is saying to use USE_AVX_INSTRUCTIONS flag enabled while dlib installation. But it is not the case for me. As I found the flag is enabled by default. I also have come through this FAQ: Why is dlib slow where the solution is to choose the Release mode in Visual Studio but I am not using Visual Studio and just running the code from terminal.

But interesting thing is that if I run the builtin face detector dlib.get_frontal_face_detector() it runs completely fine with no lag. But the program only becomes slow while running simple_object_detector() trained on custom data.

Answer 1

Alright! I got the solution. Actually I missed to provide a flag while running the detector. Probably it is related to some optimization.

First initialize the detector object:

detector = dlib.simple_object_detector("/path/to/your/trained/SVM/detector")

Now change the following line from this:

detections = detector(gray_image)

To this:

detections = detector(gray_image, 0)

Reference: https://github.com/davisking/dlib/issues/557#issuecomment-297679025