Re-distributing Biased Pseudo Labels for Semi-supervised Semantic Segmentation: A Baseline Investigation

He, Ruifei; Yang, Jihan; Qi, Xiaojuan

doi:10.1109/iccv48922.2021.00685

Cited by 80 publications

(34 citation statements)

References 28 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Consequently, the performance of the minority classes significantly degrades due to insufficient training samples. One can solve this during the training process through class re-balancing [71] or augmentation [16]. In contrast, we show that active learning can implicitly solve this by iteratively selecting a batch of samples to annotate at the data collection stage.…”

Section: Region Impurity and Prediction Uncertaintymentioning

confidence: 91%

Towards Fewer Annotations: Active Learning via Region Impurity and Prediction Uncertainty for Domain Adaptive Semantic Segmentation

Xie¹,

Yuan²,

Li³

et al. 2021

Preprint

View full text Add to dashboard Cite

Self-training has greatly facilitated domain adaptive semantic segmentation, which iteratively generates pseudo labels on the target domain and retrains the network. However, since the realistic segmentation datasets are highly imbalanced, target pseudo labels are typically biased to the majority classes and basically noisy, leading to an errorprone and sub-optimal model. To address this issue, we propose a region-based active learning approach for semantic segmentation under a domain shift, aiming to automatically query a small partition of image regions to be labeled while maximizing segmentation performance. Our algorithm, Active Learning via Region Impurity and Prediction Uncertainty (AL-RIPU), introduces a novel acquisition strategy characterizing the spatial adjacency of image regions along with the prediction confidence. We show that the proposed region-based selection strategy makes more efficient use of a limited budget than image-based or pointbased counterparts. Meanwhile, we enforce local prediction consistency between a pixel and its nearest neighbor on a source image. Further, we develop a negative learning loss to enhance the discriminative representation learning on the target domain. Extensive experiments demonstrate that our method only requires very few annotations to almost reach the supervised performance and substantially outperforms state-of-the-art methods.* Corresponding author (a) Target image (b) Image-based selection (100%) (c) Point-based selection (2.2%) (d) Region-based selection (2.2%)

show abstract

Section: Region Impurity and Prediction Uncertaintymentioning

confidence: 91%

Towards Fewer Annotations: Active Learning via Region Impurity and Prediction Uncertainty for Domain Adaptive Semantic Segmentation

Xie¹,

Yuan²,

Li³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…For example, CReST [64] selects pseudo labels more frequently for minority classes according to the estimated class distribution. DARS [20] employs adaptive threshold to select more pseudo labels for minority class during selftraining. However, these methods tends to generate noisy pseudo labels from class-biased segmentation of unlabeled data.…”

Section: Class-imbalance Learningmentioning

confidence: 99%

“…Given images X l ⊂ R H×W ×3 with pixel-level semantic labels ŷ ⊂ (1, C) H×W and unlabelled images X u ⊂ R H×W ×3 (H, W and C denote image height, image width and class number, respectively), the goal is to learn a segmentation model F that can fit both labelled and unlabelled data and work well on unseen images. Existing methods [10,20,29,31,35,43,45,47,73] combine supervised learning on labelled images and unsupervised learning on unlabelled image to tackle the semi-supervised challenge. For labelled images, they adopt cross entropy loss as supervised loss L s to train F .…”

Section: Problem Definitionmentioning

confidence: 99%

“…Semi-supervised semantic segmentation, which aims to learn from a small amount of labelled images and a large amount of unlabelled images, has been attracting increasing attention for addressing the image annotation challenge. Most existing studies tackle the challenge of semisupervised semantic segmentation by applying either consistency-training [35,46,47] or self-training [10,20,29,43,45,68] to the unlabelled data. However, they often suffer from constrained segmentation accuracy largely due to the segmentation model that is trained by using the labelled data.…”

Section: Introductionmentioning

confidence: 99%

“…This leads to class-biased segmentation of the unlabelled data which accumulates and finally degrades the whole semi-supervised learning. Though a few studies [20,64] attempt to handle the class imbalance issue by selecting more pseudo labels for minority classes during self-training, these pseudo labels are often noisy as they are generated from class-biased segmentation. Note the class imbalance issue has been widely studied in supervised learning via re-sampling [5,6,34,55,62], re-weighting [12,22,38,49] and meta-learning [49,53,63], but these works require labels to rectify biased predictions and are thus inapplicable to the unlabelled data in semisupervised semantic segmentation.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Unbiased Subclass Regularization for Semi-Supervised Semantic Segmentation

Guan¹,

Huang²,

Xiao³

et al. 2022

Preprint

View full text Add to dashboard Cite

Semi-supervised semantic segmentation learns from small amounts of labelled images and large amounts of unlabelled images, which has witnessed impressive progress with the recent advance of deep neural networks. However, it often suffers from severe class-bias problem while exploring the unlabelled images, largely due to the clear pixel-wise class imbalance in the labelled images. This paper presents an unbiased subclass regularization network (USRN) that alleviates the class imbalance issue by learning class-unbiased segmentation from balanced subclass distributions. We build the balanced subclass distributions by clustering pixels of each original class into multiple subclasses of similar sizes, which provide classbalanced pseudo supervision to regularize the class-biased segmentation. In addition, we design an entropy-based gate mechanism to coordinate learning between the original classes and the clustered subclasses which facilitates subclass regularization effectively by suppressing unconfident subclass predictions. Extensive experiments over multiple public benchmarks show that USRN achieves superior performance as compared with the state-of-the-art.

show abstract

Disease Area Detection for Chest X‐Ray Image Diagnosis Using Deep Learning with Pseudo Labeling and Ensemble Learning

Gerdprasert,

Mabu,

Kido

2023

IEEJ Transactions Elec Engng

View full text Add to dashboard Cite

The amount of labeled samples always plays a significant role in deep learning. When there are insufficient training samples, deep learning usually struggles to carry out its task successfully. The problem becomes more prominent in that labeled data are expensive to obtain. Therefore, we want to address the problem by introducing a method that utilizes unlabeled samples to improve object detection performance. We propose a framework that utilizes a pseudo‐labeling approach in semi‐supervised learning and applies it to disease diagnosis of chest X‐ray images. Our objective is to train a model that works as a labeling expert and perform the labeling task for improving the object (disease area) detection model by these newly labeled samples. In addition, we also apply an ensemble technique to our framework to reduce the model bias and improve the generalization ability for pseudo‐labeling, which results in improvements in object detection performance. Our proposed method improves the mean average precision up to 5.32, compared with the method without the pseudo‐labeling. Additionally, the mean average precision further increases by up to 19.11 when applying ensemble learning. © 2023 Institute of Electrical Engineer of Japan and Wiley Periodicals LLC.

show abstract

Re-distributing Biased Pseudo Labels for Semi-supervised Semantic Segmentation: A Baseline Investigation

Cited by 80 publications

References 28 publications

Towards Fewer Annotations: Active Learning via Region Impurity and Prediction Uncertainty for Domain Adaptive Semantic Segmentation

Towards Fewer Annotations: Active Learning via Region Impurity and Prediction Uncertainty for Domain Adaptive Semantic Segmentation

Unbiased Subclass Regularization for Semi-Supervised Semantic Segmentation

Disease Area Detection for Chest X‐Ray Image Diagnosis Using Deep Learning with Pseudo Labeling and Ensemble Learning

Contact Info

Product

Resources

About