What dataset to use for People Detection with HOG & SVM

I'm trying to do people detection with HOG & SVM algorithm. I could get maximum 60% accuracy. As I changed the training dataset, the accuracy changes but still can't pass over the %60. You think, is the problem about the dataset or the algorithm?

You can see the code in the following link: https://github.com/mfurkannargul/people-detection-hog-svm-matlab

Answer 1

[Pranav Verma]

Hi Muhammet,

The accuracy of the model trained depends on a lot of factors.

• The accuracy of the implementation of the model
• Number of samples in the dataset (both positive and negative samples)
• Correctness of the dataset ie, whether the dataset contains the objects that are expected to be classified by the model
• Kernel function being used while training the model

You can consider using Mdl = fitcsvm(___,Name,Value) where you can define options for kernel and other parameters as name, value pairs.

Similarly, you can use [___] = extractHOGFeatures(___,Name,Value) for defining options as name, value pairs.

For further information on SVM and HOG in matlab, please refer to the following documentation:

https://www.mathworks.com/help/stats/fitcsvm.html#d120e285212

https://www.mathworks.com/help/vision/ref/extracthogfeatures.html#namevaluepairarguments

Thanks