Abstract:This paper focuses on the unsupervised domain adaptation of transferring the knowledge from the source domain to the target domain in the context of semantic segmentation. Existing approaches usually regard the pseudo label as the ground truth to fully exploit the unlabeled target-domain data. Yet the pseudo labels of the target-domain data are usually predicted by the model trained on the source domain. Thus, the generated labels inevitably contain the incorrect prediction due to the discrepancy between the training domain and the test domain, which could be transferred to the final adapted model and largely compromises the training process. To overcome the problem, this paper proposes to explicitly estimate the prediction uncertainty during training to rectify the pseudo label learning for unsupervised semantic segmentation adaptation. Given the input image, the model outputs the semantic segmentation prediction as well as the uncertainty of the prediction. Specifically, we model the uncertainty via the prediction variance and involve the uncertainty into the optimization objective. To verify the effectiveness of the proposed method, we evaluate the proposed method on two prevalent synthetic-to-real semantic segmentation benchmarks, i.e., GTA5 \(\rightarrow \) Cityscapes and SYNTHIA \(\rightarrow \) Cityscapes, as well as one cross-city benchmark, i.e., Cityscapes \(\rightarrow \) Oxford RobotCar. We demonstrate through extensive experiments that the proposed approach (1) dynamically sets different confidence thresholds according to the prediction variance, (2) rectifies the learning from noisy pseudo labels, and (3) achieves significant improvements over the conventional pseudo label learning and yields competitive performance on all three benchmarks.

Mitigating Distributional Shift in Semantic Segmentation via Uncertainty Estimation From Unlabeled Data

Mitigating Distributional Shift in Semantic Segmentation via Uncertainty Estimation from Unlabelled Data

In Defense Of Multi-Source Omni-Supervised Efficient Convnet For Robust Semantic Segmentation In Heterogeneous Unseen Domains

Automated Detection of Label Errors in Semantic Segmentation Datasets via Deep Learning and Uncertainty Quantification

Hyperbolic Uncertainty Aware Semantic Segmentation

Trusting Semantic Segmentation Networks

Semantic Segmentation of Weakly Annotated Remote Sensing Images Based on Feature Adversary and Uncertainty Perception

Rectifying Pseudo Label Learning via Uncertainty Estimation for Domain Adaptive Semantic Segmentation

Estimating Uncertainty in Landslide Segmentation Models

Masked Gamma-SSL: Learning Uncertainty Estimation via Masked Image Modeling

Uncertainty estimates for semantic segmentation: providing enhanced reliability for automated motor claims handling

Uncertainty Estimation and Out-of-Distribution Detection for LiDAR Scene Semantic Segmentation

Detecting Adversarial Attacks in Semantic Segmentation via Uncertainty Estimation: A Deep Analysis

On the Calibration of Uncertainty Estimation in LiDAR-based Semantic Segmentation

Automatic Discovery and Assessment of Interpretable Systematic Errors in Semantic Segmentation

Reliability in Semantic Segmentation: Can We Use Synthetic Data?

Give me (un)certainty -- An exploration of parameters that affect segmentation uncertainty

Learning to segment subcortical structures from noisy annotations with a novel uncertainty-reliability aware learning framework

UGainS: Uncertainty Guided Anomaly Instance Segmentation

Tackling Ambiguity from Perspective of Uncertainty Inference and Affinity Diversification for Weakly Supervised Semantic Segmentation