Project-Site-Variation-Disentanglement-Siamese-Net

This is a onging research.

1. Introduction:

The purpose of this research is to design a semi-supervised network that disentangles the label-related and non-label related(variance) representation of a manifold. The label related representation denotes F_i which only has label related information(such as identity of human face). Another representation F_v contains all non-label information(such as illumination, pose and background information etc.) The F_i and F_v are very usful in:

Recognition or Clustering task such as:

• Pose, illumination, etc. Invariance Face recognition.
  
• Cross-Domain recognition
  
• Better clustering result in dataset with large variance.
  

Variance Transformation

• Stitching F_i and F_v from different images, we can transfer many appearance attributes (style, background, pose, etc.) in one image to another.

2. Task Formulation:

3. Model:

3. Clustering of F_i and F_v on MNIST:

4. Reconstruction(two image of same digit) and Cross Stitching(two image of different digit) Result on MNIST:

4. Latent Space Interpolation Result on MNIST: