Harmonizing class uniformity and separability for transferability estimation

Yuhe Ding

doi:10.1016/j.patcog.2026.113897

Harmonizing class uniformity and separability for transferability estimation

2026-05-05

0

PRE

AI

DOI：10.1016/j.patcog.2026.113897

Original

Original request for help

Save

Abstract

En 中文

• We propose HOCUS, a simple yet effective transferability estimation framework that jointly considers class separability and class uniformity. • We introduce a novel class uniformity (CU) score, defined as the entropy of the class distribution overlapping matrix, which can be flexibly integrated into existing methods as a plug-and-play metric. • Extensive experiments on image classification and semantic segmentation across diverse model zoos demonstrate that HOCUS achieves state-of-the-art transferability estimation performance.

Keywords:

transferability estimation

class uniformity

class separability

image classification

semantic segmentation

Harmonizing class uniformity and separability for transferability estimation

Abstract

Journal

Organization

Cited Papers