New paper has been accepted for presentation at the European Conference on Artificial Intelligence (ECAI2024) in Santiago de Compostela
![](/fileadmin/_processed_/9/6/csm_ACM_AI_LAB_picture22_aa7acc03c1.png)
In one of the first studies on texture and shape biases in semantic segmentation, we explore the biases of both CNNs and transformer architectures. Our approach utilizes the PDE-based anisotropic diffusion technique, termed Edge Enhancing Diffusion, originally developed by Weickert et al. in the late 90s. This technique effectively removes texture from images while preserving their shapes and allows us to study texture dependence of semantic segmentation networks.
It also turns out that training on texture-reduced data significantly enhances in-domain robustness.
The preprint is available under arxiv.org/abs/2402.09530.