ECCV2024 paper reading - Robust Nearest Neighbors for Source-Free Domain Adaptation under Class Distribution Shift
Date:
The Japanese Study Group on Computer Vision is a semestral conference in which researchers from all over Japan present the most recent topics in computer vision. In my talk, I introduce our research accepted at the European Conference on Computer Vision (ECCV2024) that addresses the problem of class distribution shift (different number of samples per class on source and target) in a source-free domain adaptation setting. Since no labels nor source data is available, it is impossible to determine the class distributions, which heavily hinders the pseudolabeling (i.e., via nearest neighbors). We propose leveraging an auxiliary model that provides a second opinion for pseudolabeling.