Enhancing Pseudo-Labeling Performance in Object Detection Using Gaussian Mixture Modeled Uncertainty

Object detection research has been rapidly advancing. However, it requires large amounts of training data, where labeling massive datasets incurs great cost and time. To address this problem, semi-supervised learning techniques have been increasingly explored, among which pseudo-labeling has become popular due to its straightforward approach. However, pseudo-labeling has limitations with confidence score-based filtering. In this paper, we propose a method to extract uncertainties using Gaussian mixture models and effectively incorporate them into the labeling process to overcome these limitations. The proposed method achieves more reliable pseudo-labeling results and experiments show a 0.8% performance improvement compared to the existing approach.

I. Introduction

The advances in deep learning have significantly improved the ability to detect complex and diverse images in computer vision tasks [1]–[3]. However, achieving high performance requires not only an enormous number of parameters and computations [4], [5], but also large-scale training data, where accurate labeling demands substantial cost and time. Motivated by this limitation, semi-supervised learning (SSL), which leverages unlabeled data to improve performance, has gained much attention [6]–[8]. In SSL research, the pseudo-labeling technique is a method that generates predicted labels for unlabeled data based on the confidence scores of a pre-trained model and uses them for training. This technique has overcome the limitations of supervised learning by simply implementing the core ideas of SSL. However, using prediction results directly as labels causes significant reliability issues. In this paper, we introduce Gaussian mixture models (GMMs) to solve this problem and extract two uncertainties (i.e., epis-temic, aleatoric) about the object. In addition, by using these uncertainties as criteria for selecting pseudo-labels, we propose a strategy to include only data with high reliability in training. As a result, the SSL performance is improved by about 0.8% compared to existing methods using only confidence scores.