Reproduce “Cityscape-Adverse: Benchmarking Robustness of Semantic Segmentation with Realistic Scene Modifications via Diffusion-Based Image Editing” 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
Naufal Suryanto, Andro Aprila Adiputra, Ahmada Yusril Kadiptya, Thi-Thu-Huong Le, Derry Pratama, Yongsu Kim, Howon Kim
Researchers at Pusan National University developed Cityscape-Adverse, a new benchmark dataset featuring diverse adverse environmental conditions, generated using advanced diffusion-based image editing while meticulously preserving semantic labels. This benchmark helps evaluate semantic segmentation model robustness and demonstrates that training with such high-fidelity synthetic data significantly enhances model generalization to real-world challenging scenarios.
Key insights
- Diffusion-based image editing models, specifically CosXLEdit, can generate high-quality, semantically consistent synthetic data for adverse conditions at scale, overcoming traditional data scarcity challenges.
- Transformer-based semantic segmentation models generally exhibit greater resilience and adaptability to out-of-distribution environmental shifts compared to CNN-based architectures.
- Training with high-fidelity synthetic adverse data from Cityscape-Adverse significantly boosts model robustness and generalization to unseen real-world challenging environments, outperforming previous synthetic data generation methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.