How to use TokyoTechFes2015

Introduction

I will explain how to use TokyoTechFes2015. I hope it will be helpful for you to practice the demo and improve the script.

environment

• Mac OS 10.11.6 (El Capitan)
• Python 2.7.10

What the script does

Face image category discrimination is performed by machine learning. By learning by changing the dataset folder and below, any face image category can be determined.

Introduction

First, clone from git.

git clone https://github.com/Arudori5001/TokyoTechFes2015.git
cd TokyoTechFes2015

• ImageMagick It is an application to take a picture with a PC camera. It is available at http://cactuslab.com/imagemagick/ (Mac).

• OpenCV http://qiita.com/zabeth129/items/f065f79b1322a82edb1d Please try to install by referring to.

• python dependent library All are available on pip. Please note that the version of chainer is quite old.

pip install cv2==1.0
pip install pillow==2.9.7
pip install chainer==1.3.2
pip install matplotlib==1.4.3

Next, unzip dataset.zip and you're ready to go.

unzip dataset.zip

The directory structure of the dataset folder created by decompressing is as follows (only the folder is displayed).

$ tree dataset -d
dataset
├── Baysters
├── Buffealoes
├── Carp
├── Dragons
├── Eagles
├── Fighters
├── Giants
├── Hawks
├── Lions
├── Marines
├── Swallows
└── Tigers

I think the names of the 12 baseball teams are lined up. Each folder contains facial images of the players of each team. If you make a prediction with the default settings, it will predict "Which team of the 12 teams the face in the given image is close to the face of the player". What you expect can be changed by changing the dataset folder and below (see below).

Forecast

1. First, take a photo of your face with ImageMagick and move the resulting image to an appropriate path.
2. Enter the following command.

python bbteam_pred.py (path of the image you want to predict)

3. The captured image will appear on the screen. If the face can be recognized, a frame will be displayed around the face. If not, the script will exit with cannot detect!.
4. If face recognition is possible, close the image in the displayed frame with the x button on the upper left, and two new screens will appear. One is an image with the text "You belong to (category)!" Written on the face of the image taken, and the other is a bar graph image. The former outputs which category was predicted, and the latter outputs the certainty (posterior probability) of each category.

Example

Here, as an example, suppose that the image ʻeagles_koyama.jpg already in the sample` folder is the image taken in step 1, and you want to make a prediction. By the way, this person is Koyama of Rakuten Eagles as the file name suggests.

2. Enter the following command.

python bbteam_pred.py sample/eagles_koyama.jpg

3. The captured image will appear on the screen. There is a frame around the face.

4. When you close the image of the displayed frame with the X button on the upper left, two new screens will appear (one may be hidden in the other window).

From the image above, we can see that Koyama is correctly predicted to be close to the face of the Eagles player, with a 95% certainty. The image below also shows confidence in teams other than Eagles. Besides the Eagles being 95%, the Carp seems to be about 5% confident.

By the way, the same output is also output to the console.

predicted categoriy : Eagles
certainty factor of predicted categoriy : 0.954637289047
certainty factors of each categories : 
Baysters :
	 0.00%
Buffealoes :
	 0.00%
Carp :
	 4.51%
Dragons :
	 0.00%
Eagles :
	95.46%
Fighters :
	 0.00%
Giants :
	 0.00%
Hawks :
	 0.02%
Lions :
	 0.00%
Marines :
	 0.00%
Swallows :
	 0.00%
Tigers :
	 0.00%

Learning

Data set changes

If you want to change what you expect, you need to change the dataset before training. The dataset uses files with a fixed directory structure in the dataset folder for learning and testing. In the default case, as mentioned above, the directory structure is as follows.

$ tree dataset -d
dataset
├── Baysters
├── Buffealoes
├── Carp
├── Dragons
├── Eagles
├── Fighters
├── Giants
├── Hawks
├── Lions
├── Marines
├── Swallows
└── Tigers

The name of each directory directly under the dataset folder (Baysters, Buffealoes, ...) represents the category name. By default, Baysters has Baystars in one category.

Images are lined up directly under folders such as Baysters.

$ tree dataset/Baysters 
dataset/Baysters
├── 1000003_ss.jpg
├── 1000007_ss.jpg
├── 1000117_ss.jpg
├── 1000118_ss.jpg
├── 1000119_ss.jpg
...

These images are face images that belong to their respective categories. Face images of Baystars players are lined up in Baysters. Please change the directory structure referring to the above.

For example, if you want to judge whether it is a man or a woman, do as follows. Save the female face image in Female and the male face image in Male.

$ tree dataset -d
dataset
├── Female
└── Male

Execution of learning script

Learning is done with the following command.

python Learning.py (number of data used for learning)(Number of data used for testing)

By default, there are 680 images in the dataset. If you want to use 520 sheets for learning and 120 sheets for testing, do as follows.

python Learning.py 520 130