3D Reconstruction

16-824 Visual Learning and Recognition: Homework 2 · Spring 2023

GITHUB

About The Project

Implemented and trained Generative Adversarial Networks (GAN) on the CUB 2011 Dataset. The primary goal was to implement GANs following provided instructions, with an emphasis on achieving specific Final FID (Fréchet Inception Distance) scores for different GAN variants: Vanilla GAN, LS-GAN, and WGAN-GP.

Built With

  • Python
  • NumPy
  • Pytorch

Results

  • Vanilla GAN:

Final FID 71.95487635262793

image

  • LS-GAN:

Final FID 89.60508651795311

image

  • WGAN-GP:

Final FID 68.6257072174157

image

  • Autoencoder image

  • VAE image

  • Diffusion

DDPM FID: 31.350106509258126

DDIM FID: 35.584719273184627 image