Insiders Hub

UnBFace

Rafaela Sinhoroto

Rafaela Sinhoroto

Brasília, Federal District

0 0
  • 0 Collaborators

A Convolutional Neural Network based on InceptionV3 architecture trained to estimate the face pose (described in yaw, pitch and roll angles) from a digital, RGB image of the user's head, for control system applications. ...learn more

Project status: Under Development

Virtual Reality, Robotics, Artificial Intelligence

Intel Technologies
DevCloud, Intel Python

Code Samples [1]Links [1]

Overview / Usage

The main objective of this project is to extract information that can be used as control system input (for example, a motorized wheelchairs driving system, auxiliar robotic arms for people with reduced mobility, among others). A Convolutional Neural Network based on the InceptionV3 architecture is used to estimate the angles that describe the head pose orientation of the subject in tridimensional space from a digital RGB picture as input.

Methodology / Approach

The network is implemented in Python v3.6 using the Keras API running on the TensorFlow framework; When the network is trained using Google Colaboratory, the regular TF implementation is used; When it's on DevCloud environment, the intel optimized version of TF is installed.

The datasets chosen for training are the AFLW, the 300W-LP and the BIWI Kinect Head Pose dataset.

The process consists of training the network multiple times with varying weight initializations with the goal to find a good initialization that allows for a better minimization of the loss function (mean squared error). The performance of the network is measured using Mean Absolute Error and comparing with other similar published implementations (such as Hsu et al. QuatNet).

Technologies Used

  • Keras
  • TensorFlow
  • Python v3.6
    • scikit-image
    • scikit-learn
  • Google Colaboratory
  • Intel DevCloud

Repository

https://github.com/RSinhoroto/UnBFace

Comments (0)