A.Project Description

Project Details

The project detail is in 6000C Homepage.

Due to the intranet’s limitation, you can click the Brown CS143 for temporary reference.

Develop environment

Develop Environment: MATLAB + vl_feat

Programming language: MATLAB

System: MacOS 10.13

B.Final result

1. training performance

C.Steps

1. Load positive training crops and random negative examples

details of load positive training crops：

extract file names of images
load the images
Rgb2gray operation and Normalization
compute HOG descriptors
save the HOG features

Details of load random negative examples:

compute the candidates of possible values of x_min,y_min of cropped images.
parallel implement the generating candidates of negative samples.
extract their HOG features and save it.

2. Train Classifier

Details of training:

concatenate the features of positive and random negative samples along the x-axis. And transpose the whole X matrix(feature matrix).
concatenate the labels of positive and negative samples.
set the lambda(The regularization parameter).
train the w and b by vl_svmtrain.

3. Examine learned classifier

visualize the perfect HOG descriptors trained for human detection.

Visualize how well separated the positive and negative examples are at training time.

4. Run detector on test set.

Details

set the pre-filtering threshold.
for a single image, scale the original image by various times.
compute all candidates of possible x_min,y_min,x_max,y_max.
compute the corresponding boxes in original image.
send all candidates filtered by threshold into non-max suppression function.
save the possible result for face detection.

5. Evaluate and Visualize detections

Match the figure 6 Average precision

Sample result

D. Bonus

1. Hard negative mining