Validation of annotation reduction in pose estimation models using semi-supervised learning

Abstract

This paper aims to analyze the effects of pseudo labels for training pose estimation models using semi-supervised learning. One of the problems in training pose estimation models is the high cost of labeling, which requires keypoint annotation of training data. Pseudo-labels estimated by other pose estimator models are able to improve the performance of another estimator. This paper investigates the trade-off between the number of annotations by humans and pseudo labels given by the model.