A Three-Stage Self-Training Framework for Semi-Supervised Semantic Segmentation

Ke, Rihuan; Avilés-Rivero, Angelica I.; Pandey, Saurabh; Reddy, Saikumar; Schönlieb, Carola‐Bibiane

doi:10.1109/tip.2022.3144036

Cited by 37 publications

(8 citation statements)

References 44 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The TPLD method [35] uses a twostage self-training framework, first exploring the easier parts of the target domain images, and then processing the more difficult parts. TSST [36] proposes a three-stage self-training framework, which uses multiple stages of multi-task learning strategies to constrain and generate better pseudo-labels. RoadDA [37] uses a similar approach to TPLD to study road segmentation in remote sensing images.…”

Section: Stage-wise Self-trainingmentioning

confidence: 99%

Semantic Segmentation of Foggy Scenes Based on Progressive Domain Gap Decoupling

Wang¹,

Zhang²,

Ma³

et al. 2023

Preprint

View full text Add to dashboard Cite

Robust Semantic Foggy Scene Segmentation (SFSS) is crucial for the security of autonomous driving. However, the blurring of images caused by fog increases the difficulty of recognition and makes annotation of foggy scene images more expensive, resulting poor performance when recognizing entities in the fog. Currently, many methods use domain adaptation to transfer segmentation knowledge from clear scenes to foggy ones. But these method are often ineffective due to the large domain gap between different cities' styles and the quality degradation of images caused by fog. The latest research has attempted to introduce an intermediate domain to decouple the domain gap and gradually complete the semantic segmentation of foggy scenes, but the exploration of the intermediate domain information is often insufficient. To solve these problems, we first analyze the self-training in domain adaptation and propose the concept of "label reference value". We prove that the higher the total label reference value, the easier the self-training performance gets improved. With this precondition, we can reasonably split the original problem into two-stage domain adaptation. In each stage, the "label reference value" can be controlled and maximized. Specifically, the first stage only process the style gap between source domain and the intermediate domain, and the second stage process the fog gap. The fog gap includes: (1) real fog gap between the intermediate domain and target domain; (2) the synthetic fog gap between clear source domain and synthetic foggy source domain. This allows the model to make full use of "label reference value" and gradually develop good semantic segmentation skills for foggy scenes. Our approach significantly outperforms the baseline algorithm on all the mainstream SFSS benchmarks, with good generalization ability demonstrated in other adverse scenes such as rain and snow. We also compare our method with latest large segmentation models, which shows that our method has more robust performance in the foggy scenes.

show abstract

Section: Stage-wise Self-trainingmentioning

confidence: 99%

Semantic Segmentation of Foggy Scenes Based on Progressive Domain Gap Decoupling

Wang¹,

Zhang²,

Ma³

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…Current deep SSL classification methods utilize entropy minimization and consistency regularization to learn the unlabeled data. Entropy minimization based methods enforce the predictions of unlabeled data to be sharp which is commonly realized by training their model via pseudo labels [2], [3], [20], [24], [39], [49]. Consistency regularization based methods enforce their models to predict consistent results for the same image in different perturbations to make the classification boundaries pass through the data sparse region.…”

Section: A Semi-supervised Learningmentioning

confidence: 99%

PCR: Pessimistic Consistency Regularization for Semi-Supervised Segmentation

Qiao¹,

Wei²,

Wang³

et al. 2022

Preprint

View full text Add to dashboard Cite

Currently, state-of-the-art semi-supervised learning (SSL) segmentation methods employ pseudo labels to train their models, which is an optimistic training manner that supposes the predicted pseudo labels are correct. However, their models will be optimized incorrectly when the above assumption does not hold. In this paper, we propose a Pessimistic Consistency Regularization (PCR) which considers a pessimistic case that pseudo labels are not always correct. PCR makes it possible for our model to learn the ground truth (GT) in pessimism by adaptively providing a candidate label set containing K proposals for each unlabeled pixel. Specifically, we propose a pessimistic consistency loss which trains our model to learn the possible GT from multiple candidate labels. In addition, we develop a candidate label proposal method to adaptively decide which pseudo labels are provided for each pixel. Our method is easy to implement and could be applied to existing baselines without changing their frameworks. Theoretical analysis and experiments on various benchmarks demonstrate the superiority of our approach to state-of-the-art alternatives.

show abstract

“…This teacher model assigns pseudo-labels to selected unlabeled nodes and incorporates them to augment the labeled data, after which a student model is trained on the augmented training nodes. Typically, current self-training frameworks introduce multiple stages where the teacher model is iteratively updated with augmented data to generate more confident pseudo-labeled nodes [7,11,29,37]. For example, Self-Training (ST) [16] represents a single-stage self-training framework that initially trains a GCN using given labels, subsequently selects predictions exhibiting the highest confidence, and then proceeds to further train the GCN utilizing the expanded set of labels.…”

Section: Introductionmentioning

confidence: 99%

Distribution Consistency based Self-Training for Graph Neural Networks with Sparse Labels

Wang,

Zhao,

Wang

2024

Proceedings of the 17th ACM International Conference on Web Search and Data Mining

View full text Add to dashboard Cite

Few-shot node classification poses a significant challenge for Graph Neural Networks (GNNs) due to insufficient supervision and potential distribution shifts between labeled and unlabeled nodes. Self-training has emerged as a widely popular framework to leverage the abundance of unlabeled data, which expands the training set by assigning pseudo-labels to selected unlabeled nodes. Efforts have been made to develop various selection strategies based on confidence, information gain, etc. However, none of these methods takes into account the distribution shift between the training and testing node sets. The pseudo-labeling step may amplify this shift and even introduce new ones, hindering the effectiveness of self-training. Therefore, in this work, we explore the potential of explicitly bridging the distribution shift between the expanded training set and test set during self-training. To this end, we propose a novel Distribution-Consistent Graph Self-Training (DC-GST) framework to identify pseudo-labeled nodes that both are informative and capable of redeeming the distribution discrepancy and formulate it as a differentiable optimization task. A distributionshift-aware edge predictor is further adopted to augment the graph and increase the model's generalizability in assigning pseudo labels. We evaluate our proposed method on four publicly available benchmark datasets and extensive experiments demonstrate that our framework consistently outperforms state-of-the-art baselines. CCS CONCEPTS• Computing methodologies → Semi-supervised learning settings; Neural networks.

show abstract

A Three-Stage Self-Training Framework for Semi-Supervised Semantic Segmentation

Cited by 37 publications

References 44 publications

Semantic Segmentation of Foggy Scenes Based on Progressive Domain Gap Decoupling

Semantic Segmentation of Foggy Scenes Based on Progressive Domain Gap Decoupling

PCR: Pessimistic Consistency Regularization for Semi-Supervised Segmentation

Distribution Consistency based Self-Training for Graph Neural Networks with Sparse Labels

Contact Info

Product

Resources

About