Bilkent University
Department of Computer Engineering

---

CS 590/690 SEMINAR

 

Optimally Matched Hierarchy for Semantic Segmentation

 

Ali Baran Özaydın
Master Student
(Supervisor:Asst.Prof.Doruk Öner)

Computer Engineering Department
Bilkent University

Abstract: Existing Unsupervised Semantic Segmentation (USS) methods fail to optimize feature projection and clustering objectives jointly. This leads to segments that are often misaligned with the clustering goals due to ambiguous class definitions. We propose a new differentiable paradigm that unites these processes. Our method learns a soft hierarchy among parallel cluster levels via Optimal Transport and optimizes the feature projection head towards aligned clustering levels. This enables features to capture multi-granularity information, improving alignment between features and segments. Our experiments demonstrate that optimizing the feature projection head and clustering head jointly boosts the performance on USS on multiple datasets and models.

 

DATE: November 17, Monday @ 15:50 Place: EA 502