The paper presents ProLab, a unique approach that uses property-level label space to generate robust and interpretable segmentation models. Rather than solely relying on category-specific annotations, ProLab uses descriptive properties grounded in common sense knowledge for supervising segmentation models. This is achieved by employing Large Language Models (LLMs) and carefully crafted prompts to generate descriptions of all involved categories. These descriptions are then embedded and clustered into a set of descriptive properties using K-Means. The paper shows that ProLab enhances the performance of segmentation models on five classic benchmarks and demonstrates better scalability with extended training steps.
Publication date: 21 Dec 2023
Project Page: https://github.com/lambert-x/ProLab
Paper: https://arxiv.org/pdf/2312.13764
