Reproduce “Image Translation with Kernel Prediction Networks for Semantic Segmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Cristina Mata, Michael S. Ryoo, Henrik Turbell
DA-KPN introduces a lightweight, Kernel Prediction Network-based method for unpaired image translation that explicitly preserves semantic structure during domain adaptation for dense prediction tasks. It generates photorealistic images from synthetic inputs while maintaining accurate label integrity, resulting in improved semantic segmentation performance on benchmarks like GTA→Cityscapes.
Key insights
- Explicitly modeling domain differences through a fixed, symmetric pixel-wise transformation function guarantees geometric preservation, fundamentally preventing semantic flipping during image translation.
- Guiding the translation process with stable semantic features from a frozen encoder enables interpretable and effective domain adaptation without requiring paired data.
- A computationally efficient KPN architecture, combined with multi-scale adversarial training, provides a scalable solution for high-resolution images while outperforming GANs in downstream task performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.