Thanks for joining us for SIGGRAPH 2020. Will we see you at SIGGRAPH 2021?

Poster
:
Reflection Removal via Realistic Training Data Generation
Event Type
Poster
Interest Areas
Arts & Design
Research & Education
Registration Levels
US
UA
EN
EU
Keywords
AI/Machine Learning
Rendering
Time
LocationCarousel
DescriptionWe propose a valid polarization-based reflection contaminated image synthesis method, which provides an adequate, diverse, and authentic training dataset, enhances the neural network by introducing the reflection information as guidance, and utilizes adaptive convolution kernel size to fuse multiscale information, which significantly improves the single image reflection removal results.