Semantic image segmentation via dynamic curriculum learning

Xiang Zhang, Wanqing Zhao, Chenji Wang, Hangzai Luo, Sheng Zhong, Lei Tang, Jinye Peng, Jianping Fan

Abstract: Manually annotating pixel-level labels for large-scale data in semantic segmentation tasks is not only time-consuming and laborious but also challenging to ensure consistent annotation quality. In this paper, we present a simple yet effective semantic segmentation framework, named Dynamic Curriculum Learning (DCL), to progressively learn the segmentation knowledge by harnessing synthetic and unlabeled images. Initially, a data generation strategy is proposed to generate large-scale synthetic images with pixel-level labels automatically. Subsequent iterations involve a dual optimization process wherein the model is refined using both the synthetic images with their inherent pixel-level labels and the target images augmented with the generated pseudo labels. After the training of each round, DCL dynamically revises the training dataset and the pseudo labels for the target domain through a tripartite strategy encompassing Image-level Sample Selection (ISS), image augmentation, and Conditional Random Fields (CRF) post-processing. Without using any labels (i.e., image-level labels, bounding-box labels, and pixel-level labels) of target data, our model yields state-of-the-art performance than most existing semi-supervised and synthetic image approaches.

External IDs:dblp:journals/apin/ZhangZWLZTPF25