From 6d61486f13342d67a0d28060e050ebee1bbc262d Mon Sep 17 00:00:00 2001 From: americast Date: Fri, 9 Mar 2018 22:06:14 +0530 Subject: [PATCH] Add paper on HOG --- ... Oriented Gradients for Human Detection.MD | 41 +++++++++++++++++++ 1 file changed, 41 insertions(+) create mode 100644 papers_discussed/Histograms of Oriented Gradients for Human Detection.MD diff --git a/papers_discussed/Histograms of Oriented Gradients for Human Detection.MD b/papers_discussed/Histograms of Oriented Gradients for Human Detection.MD new file mode 100644 index 0000000..fc58bb0 --- /dev/null +++ b/papers_discussed/Histograms of Oriented Gradients for Human Detection.MD @@ -0,0 +1,41 @@ +# Title + +Histograms of Oriented Gradients for Human Detection + +# tl;dr + +Making use of Histograms of Oriented Gradients for detecting humans in images. Makes use of Gamma normalisation, gradient computation (using a simple [-1, 0, 1] filter), making histograms having 9 bins, from 0 to 180 degress, generating feature vector and performing contrast normalisation on top of it. The final feature vector is classified in a binary fashion using linear SVM. + +# Describe the method + +## Architecture +An image is processed in the following manner. + +### Gamma normalisation +Square root operation is performed on the image, followed by a minmax normalisation + +### Gradient computation + +Gradient is computed horizontally and vertically using a [-1, 0, 1] filter and it's transposed. Again, this is followed by a minmax normalisation + +### Binning + +HOGs are genrated by placing angles in bins. No. of bins is set to 9, each 20 degrees apart. The paper instructs moving the values obtained at the edges to next bins, and this is achieved via bilinear interpolation. + +### Contrast normalisation + +The feature vectors are contrast normalised. L2-Hys, L2-norm and L1-sqrt +worked best using Bhattacharya distance. 64×128 detection window includes about 16 pixels of margin around the person on all four sides. + +Then, classification is performed in a binary fashion using linear SVM. + +## Dataset +MIT pedestrian database and INRIA dataset + +# Any further details + +[Bhattacharya Distance](https://en.wikipedia.org/wiki/Bhattacharyya_distance) + +# My two cents + +Making histograms from 0 - 360 degrees in place of 0 - 180 degrees helps making object detection algorithms instead of human detection. \ No newline at end of file