Efficient Iterative Active Learning for Semantic Segmentation Using Sparse Point Annotations

Abstract

Translate Article

This study introduces a novel iterative active learning framework for semantic segmentation of remote sensing images, designed to minimize annotation effort while achieving competitive performance. The framework consists of an integrated interface that allows point labeling and model training in a single platform, along with the proposition of a new loss function, the Dynamic Weighted Confidence Dice Loss, which dynamically penalizes confident incorrect points. To highlight the method’s efficiency even under strict annotations, we labeled at most one foreground and one background pixel per frame in each iteration. The experiments were conducted on the BSB Aerial dataset, containing RGB bands with 0.24m image spatial resolution. The dataset included 1000 training, 200 validation, and 200 test frames, each 256×256 pixels, in a binary scenario (road and background). The framework was compared against dense annotations, random sparse annotations, and the standard Dice loss (with the same selected points in the iterative process). Dense annotations, which required labeling all pixels, achieved an Intersection over Union (IoU) of 89.63%, serving as an upper bound, while random sparse annotations reached an IoU of 75.20%, serving as a lower bound. By the sixth iteration, our method achieved an IoU of 86.46%, only 3.17% lower than dense annotations, while labeling just 0.0136% of the total dataset. Furthermore, the proposed loss consistently outperformed the Dice loss during iterative training, with IoU improvements ranging from +0.17% in the first iteration to +4.17% in the final iteration, highlighting that its effectiveness increases as more points in error-prone regions are annotated.