EAD-GAN: A Generative Adversarial Network for Disentangling Affine Transforms in Images

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Generative Adversarial Networks (GANs) are powerful tools for data generation.
• Disentangling factors of variation in latent spaces is crucial for controllable generation.
• Existing methods often struggle to explicitly separate affine transformations.

Purpose of the Study:

• To propose a novel GAN, the Explicit Affine Disentangled Generative Adversarial Network (EAD-GAN).
• To achieve self-supervised disentanglement of affine transformations.
• To enable generation of specifically transformed data by controlling learned parameters.

Main Methods:

• Developed an affine transform regularizer to enforce explicit affine properties in InfoGAN.
• Decomposed the affine matrix and used least-squares for parameter inference.
• Trained EAD-GAN on MNIST, CelebA, and dSprites datasets.

Main Results:

• EAD-GAN successfully disentangles affine attributes including rotation, zoom, skew, and translation.
• Learned representations possess clear physical meaning.
• Achieved superior disentanglement scores on the dSprites dataset compared to state-of-the-art methods (MIG: 0.59 vs 0.37, DCI: 0.96 vs 0.71).

Conclusions:

• EAD-GAN offers an effective approach for self-supervised disentanglement of affine transformations.
• The method learns interpretable representations with direct physical meaning.
• EAD-GAN demonstrates significant improvements in disentanglement performance over existing techniques.