Jayne Wilson

Get Negative Features

This process was similar to obtaining the positive features. The difference was that for the negative features, each image was sampled at random and each sample had its HOG computed. The HOGs from the sample were used to populate negative features descriptor matrix. I used an average number of samples per image.

Classifier Training

To train the classifier I applied vl_svmtrain to the newly found positive and negative feature sets. I started with a lambda of 0.001, then changed to 0.0001, with increased my accuracy from around 0.9997 to 0.9998 or higher. To create the data for the trainer, I combined and transposed the feature sets. Arrays of labels for the examples of positive and negative images were made of arrays of +1 and -1, respectively.

Results:

Here's what my results looked like from the initial classification:

 
Initial classifier performance on train data:
  accuracy:   1.000
  true  positive rate: 0.398
  false positive rate: 0.000
  true  negative rate: 0.602
  false negative rate: 0.000
vl_hog: descriptor: [6 x 6 x 31]
vl_hog: glyph image: [126 x 126]
vl_hog: number of orientations: 9
vl_hog: variant: UOCTTI
Detecting faces in Argentina.jpg
 non-max suppression: 619 detections to 70 final bounding boxes
 

Detector Implementation

To implement the sliding window detector, I took each normalized, grayscale image and scaled it. Then I computed the HOG features of the scaled image using vl_hog. Next I used the template dimensions to create a window equal to the template size. The number of cells per image was the number of hog cells (6). The HOG image was then computed for each window as it traced through the image, and similar to the get_features functions, it created a feature array of the computed (and vectorized) HOG descriptors. Once the descriptors were computed, their confidence was scored using the w and b values trained with my classifier. The scores for the features were then compared to the threshold and those that were above that level were chosen to compute the bounding boxes.

Here is a code snippet of the sliding window algorithm.

 
%sliding window
        for n = 1 : winx
            for m = 1 : winy
                %calculate the index of the feature
                k = (winy * (n-1) + m);
                %crop the image around the feature (for bounding box)
                winfeat = hog(m : (m + num_hog_cells - 1), n : (n + num_hog_cells - 1), :);
                %reshape to 1x(dimensionality of HOG image) add to descriptor results
                window(k, :) = reshape(winfeat, 1, (num_hog_cells^2*31));
            end
        end
		
		
Confidence scoring:

confidence_scores = window * w + b;
best = find(confidence_scores > threshold);