Domain Adaptation from Sequentially Arriving Target Images with Online Deep Sets

Abstract

This paper addresses domain adaptation of the visual inspection model from sequentially arriving images, where we start domain adaptation from the first image we observe, i.e., zero-shot adaptation. It contributes to the rapid start of automated visual inspection. We propose to encode domain information as a vector considering the sample size for the domain adaptation in this sequential setting (online domain vector). Based on the vector, we derive the loss to improve the few-shot adaptation performance. The vector and the loss are efficiently computed with online deep sets.