Taras Lehinevych "Shadows Generation in the Wild"
Download as PPTX, PDF2 likes383 views
The document provides an overview of techniques for shadow generation and removal using generative adversarial networks (GANs), specifically focusing on CycleGAN, MaskShadowGAN, and U-GAT-IT. It discusses various datasets used for shadow detection, including the SBU and ISTD datasets, and highlights the motivations behind improving shadow management in images. The document concludes with an invitation for questions regarding the methods and findings presented.
Taras Lehinevych "Shadows Generation in the Wild"
- 1. Shadows Generation in the Wild Taras Lehinevych Rails Reactor
- 2. Who am I ML Engineer @ Rails Reactor ML Engineer @ nice try bro :) (censored) Disclaimer! All the provided information based on open publication and datasets
- 3. Agenda - Motivation - Generative Adversarial Networks - Cycle GAN - MaskShadowGAN - U-GAT-IT - Summary
- 4. What’s wrong with shadows?
- 5. What do we want?
- 6. Datasets SBU Dataset (link) - this new dataset contains 4,727 images (4,089 train images and 638 test images) with pixel based ground truth. ISTD Dataset (link) - it contains 1870 triplets of shadow, shadow mask and shadow-free image under 135 different scenarios.
- 8. GAN objective Generative Adversarial Networks
- 9. Conditional GAN Image-to-Image Translation with Conditional Adversarial Networks
- 10. Conditional GAN Image-to-Image Translation with Conditional Adversarial Networks z|yx|y
- 11. Shadow Detection Fast Shadow Detection from a Single Image Using a Patched Convolutional Neural Network (2018)
- 12. Fast Shadow Detection from a Single Image Using a Patched Convolutional Neural Network (2018)
- 14. Approach #2 - Shadow Detection & Removing Stacked Conditional Generative Adversarial Networks for Jointly Learning Shadow Detection and Shadow Removal (2017)
- 16. Shadow Detection & Removing
- 17. Approach #3 - Shadow Generation/Augmentation
- 18. Cycle GAN - Datasets https://hardikbansal.github.io/CycleGANBlog/
- 19. Cycle GAN Unpaired Image-to-Image Translation using Cycle- Consistent Adversarial Networks (2017) https://hardikbansal.github.io/CycleGANBlog/
- 20. Cycle GAN
- 21. Cycle GAN
- 22. Identity Loss
- 23. Cycle GAN
- 24. Mask-Shadow GAN Mask-ShadowGAN: Learning to Remove Shadows from Unpaired Data
- 25. Mask-ShadowGAN
- 26. Mask-ShadowGAN
- 27. What to improve? Mask-ShadowGAN: Learning to Remove Shadows from Unpaired Data
- 28. U-GAT-IT U-GAT-IT: Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization for Image-to-Image Translation
- 29. U-GAT- IT
- 30. U-GAT-IT
- 31. Summary
- 32. Thanks for your time and attention! Questions?