renatobmlr/facial_keypoints_detection

Predict keypoint positions on face images

Facial Keypoints Detection

Description

The objective of this task is to predict keypoint positions on face images. The dataset used is from the Kraggle competition: https://www.kaggle.com/c/facial-keypoints-detection.

The Facial Keypoints Detection was made using PyTorch, which is a python module that uses Tensors and dynamical Neural Networks to optimize GPU operations.

Requirements

• Python 3.6+

• Jupyter Notebook 5.0+

• Numpy 1.13.1+

• Matplotlib 2.0.2+

• PyTorch

Installation

This notebook was developed in Jupyter Notebook using Anaconda environment. This is not a requirement to run the notebook provided in this repository, however as it was used to develop it therefore it and it is a well known environment among data scientists is provided a quick guide in how to install and set up your environment.

Anaconda

Go to the link and download the version suited to your platform.

The Installation process is quite simple just follow the assistant that is provided when it is launched the application.

After the installation is done open a terminal and check the version installed using the following code via terminal :

$ conda -V

Then check for the packages using the following :

$ conda list

Check for the versions of listed on Requirements sessions and in case that is something out of data except for Keras that is not a default package use the following code via terminal.

$ conda update anaconda

facial_keypoints_detection's Issues

Why the label y needs to substract 48 and divide by 48?

I want to ask that why the label y needs to substract 48 and divide by 48 like:
y = (y - 48) / 48

use self.img_original to crop instead of self.img_with_detections

Thanks for the code!
In the detection method of the FaceDetection class you might want to crop out faces on the original image instead of img_with_detections:

def detection(self):
    faces_crop = []
    for (x, y, w, h) in self.faces:  
        obj = self.img_original[y:y + h, x:x + w]
        faces_crop.append(obj)
        cv2.rectangle(self.img_with_detections, (x, y), (x+w, y+h), (0, 255, 0), 2)
        
    return faces_crop

Otherwise in the case of overlapping boxes, you get cropped images containing lines from other detection boxes.