Adversarial Scene Editing: Automatic Object Removal from Weak Supervision

Learn about a novel approach to automatic object removal in images using weak supervision in this 30-minute lecture from the University of Central Florida. Explore the two-stage process involving a mask generator and image inpainter, delving into their architectures, training methods, and loss functions. Examine the dataset used, evaluation metrics, and quantitative results, including removal performance and image quality assessment. Gain insights into potential failure cases and the findings from ablation studies in this comprehensive overview of adversarial scene editing techniques.

Intro
Goal
Contemporary problems
Contribution
Approach
Stage 1 - Mask generator GM
Stage 2 - Image inpainter G
Architecture - Mask generator
Architecture - Image inpainter
Training - Mask generator
Training - Image inpainter
Mask priors
Mask generator loss
Final loss function - Mask genera
Optimizing inpainter - local labe
Final loss function - Image inpair
Dataset - 1
Evaluation metrics
Removal performance
Image quality assessment
Human evaluation
Quantitative results
Failure cases
Ablation study - 1