Semantic Segmentation Active Learning with Scene Coverage CoreSet

Hailun Liang, Sunyuan Qiang, Hui Ma, Jun Wan, Yanyan Liang

Abstract: Semantic segmentation typically necessitates laborious pixel-level annotations for training. In response, semantic segmentation active learning (SSAL) paradigms have emerged, aiming to reduce these costs by selecting the most valuable samples for annotation. Recent SSAL methods mainly prioritize evaluating sample annotation value based on data coverage. However, the overlooked poor data coverage during the cold start phase severely limits the annotation performance. To this end, this paper introduces a novel Scene Coverage CoreSet (SC-CoreSet) method to reduce coverage loss and boost active learning performance. Specifically, the proposed method enhances image representativeness by utilizing a scene-aware Gaussian mixture with well-established latent space to ensure the selection of the most representative images. Additionally, we integrate flood regularization derived from the Pareto frontiers to reduce coverage loss. Experiments on the Cambridge-driving Labeled Video Database dataset (Camvid) validate the superior performance of our proposed method.

External IDs:dblp:conf/ccbr/LiangQMWL24