Goggle detetion with HOG + SVM + Hard Neg Training

Author: Zelong Cong ,University of Melbonrne

Date: 1st June, 2017

overview

This is an application built to detect human and identify whether he/she is wearing a pair of goggles from the image. Two models, for Human and goggle, were used in the application. I tested the Haar and the HOG descriptor using Adaboost and LinearSVM respectively, plus the model comes with Opencv3.2.0

Evironments:

1. OpenCV Module 3.2.0 (latest version) (This is a possible guide to install the module, though it not always works for some reasons... http://www.pyimagesearch.com/2016/12/05/macos-install-opencv-3-and-python-3-5/)

2. Python 3.6.2

Datasets:

For Humand Detection
in the body detection part, I used INRIA Human dataset plus some custom pics.
In the face detection part, I used AT&T Cambridge Human Face dataset.
For goggle detector, I used 200s custom made pics.

Hard Negative Mining

According t There are dozens of codes in Github to show re-creation of Dalal's human detection but none of them showed the most critical part- Hard Negative Training(refer to Dalal's). It

A list of required packages:

1. scikit-learn
2. scikit-image
3. cv2 (python wrapper for OpenCV)
4. numpy
5. imutils
6. matplotlib

What this archive contains

1.This archive contains 8 major applications created during the Project:

'visualization.py'    ---- The application used to visualise the image's HOG features.

'gather_feature.py'   ---- The application used to extract features.

'train_svm.py'        ____ The application used to train a custom made SVM model.

'hard_neg_mining.py'  ---- A very important application used to apply hard negative training process.

'detection.py'    ---- The most siginificant application achieved our goal, to detect human and identify whether he is wearing a pair of goggle with OpenCV detector.

'tool/Image_net_downloader.py'  ---- The application used to download images to local machine from ImageNet

'tool/image_resize.py'     ---- The application used to resize the images including flipping operation.

'tool/test_custom_detector.py'  ---- The application used to experiment the goggle detection.

2. The models for test locates in the directory 'models'

Guide to test

IMPORTANT: better to use absolute path to run applicatioin.

1. To use the 'goggle_detection_opencv.py'

Use command $ 'python3 goggle_detection_opencv.py -i path/to/an/image -f path/to/face/detector -g path/to/model/goggle'

for example in this directory, 

'python3 goggle_detection_opencv.py -i test_image/IMG_0518.JPG -f models/detect_goggle.xml -g models/face_detector.xml'

2. To run 'gather_feature.py'

Use command $ 'python3 gather_feature.py -p -n -f -g'
-p is followed by the path to positive images directory.
-n is followed by the path to negative images directory.
-f is followed by the path to positive features directory.
-g is followed by the path to negative features directory.

3. To run 'train_svm'

Use command $ 'python3 train_svm.py -p -n -s'
-p is followed by the path to positive features directory.
-n is followed by the path to negative features directory.
-s is followed by the name of the model.

4. To run 'hard_neg_mining.py'

Use command $ 'python3 hard_neg_mining.py -i -p -n -r -m'.
-i is followed by the path to directory contains neg samples.
-p is followed by the path to pos features.
-n is followed by the path to neg features.
-r is followed by the starting round of hard negative mining.
-m is followed by the path to the fist model used in hard negatie training.

5. To run 'visualization.py'

Use command $ 'python3 visualization.py -s'
-s is followed by the path to source directory containing the images.

6. To run 'image_resize.py'

Use command $ 'python3 image_resize.py -t -s -f '
-t is followed by the path to store the resized images
-s is followed by the path to a directory contains images to be resized
-f is followed by the option 'flip', default is 'No', give 'Yes' to enable flipping.
as a result of flip mode, any image resized will be flipped and stored again.

7. To run 'Image_net_downloader.py'
   before running this application, you have to copy all the links of your interest cateogry from ImageNet and Paste them in a 'txt' file, then change the config in python script to modified the path to txt file and the destination of output.
   

Use 'python3 Image_net_downloader.py' directly.

8. To run 'test_custom_detector.py'

Use 'python3 test_custom_detector.py -i -f -g'
-i is followed by the path to test image.
-f is the path to facial detector
-g is the path to the goggle detctor.

TESTTEST