Search:
Match:
1 results

Analysis

This paper addresses the limitations of self-supervised semantic segmentation methods, particularly their sensitivity to appearance ambiguities. It proposes a novel framework, GASeg, that leverages topological information to bridge the gap between appearance and geometry. The core innovation is the Differentiable Box-Counting (DBC) module, which extracts multi-scale topological statistics. The paper also introduces Topological Augmentation (TopoAug) to improve robustness and a multi-objective loss (GALoss) for cross-modal alignment. The focus on stable structural representations and the use of topological features is a significant contribution to the field.
Reference

GASeg achieves state-of-the-art performance on four benchmarks, including COCO-Stuff, Cityscapes, and PASCAL, validating our approach of bridging geometry and appearance via topological information.