Efficient data selection method for semi-supervised semantic segmentation

Abstract

Semantic segmentation, which provides dense information by classifying pixels into specific classes and accurately providing object boundaries in images, has gained significant attention in various fields such as autonomous driving and medical image processing. However, acquiring high-quality training data for this task can be time-consuming and expensive. To address this challenge, researchers have explored semi-supervised learning based approaches that utilize few labeled and large amounts of unlabeled data. Nevertheless, most existing methods randomly sample labeled data from the entire dataset, leading to significant variations in the performance of the trained models, particularly when labeled data is scarce.

In this paper, we propose a solution to this problem by introducing an efficient data selection method for training semi-supervised semantic segmentation networks. Our method involves selecting the required labeled data from the overall dataset, leveraging the diversity within and across different images based on widely used pre-trained models. Notably, our approach does not require human interaction or additional training for data selection. We validate the effectiveness of our proposed method through experiments conducted on the Pascal VOC and Cityscapes datasets, using various semi-supervised semantic segmentation networks. The results show that our method achieves higher and more consistent performance compared to the random data selection method.