Presented by: Elaheh Barati

elaheh@wayne.edu

Wayne State University

Multi-view CNN for Face Verification in Videos

Objective:

• Our goal is to perform video based face verification using a stream based CNN

Face

Verification

Face Verification System

Same or Not?

Video-based Face Verification

feed a video as a
sequence of frames

feed video directly into the ConvNets as an input

Video-based Face Verification

feed a video as a
sequence of frames

feed video directly into the ConvNets as an input

Combine information from a sequence of frames using a CNN architecture with stream pooling and fully connected layers.

Our approach:

Data Pre-processing:

• Face Detection
• Face Tracking

[1] Zhang, Kaipeng, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. "Joint Face Detection and Alignment Using Multitask Cascaded Convolutional Networks." IEEE Signal Processing Letters 23, no. 10 (2016): 1499-1503.

[1]

original

Positive

Negative

Original

Positive

Negative

Original

Positive

Negative

Video Face DataSet

Network Architecture

CNN1:

• Alex-Net Architecture
  • ​input size : 227 x 227
• Initialize the weights of convolutional layers with the weights from Alex-Net model which was trained on the ImageNet

Triplet Loss:

The Triplet Loss minimizes the distance between an anchor and a positive, both of which have the same identity, and maximizes the distance between the anchor and a negative of a different identity

Works to be done

• Train with more data
  • IJB-A dataset
• Video to video face verification
• Testing on 3D face images