Effect of Human-Selected Hard Examples to Improve Accuracy of Semantic Segmentation

An accurate classifier is crucial for improving the reliability of a visual navigation method based on semantic segmentation. However, it is challenging to develop a comprehensive dataset that includes all environmental conditions. Here, we selected misclassified images, termed hard examples, and augmented the existing training data to train a reliable classifier with limited additional data. Experimental results using actual images captured during the course of the Tsukuba Challenge 2024 indicated that the addition of hard examples improves the segmentation accuracy in actual scenarios.