Uncertainty-based method for improving poorly labeled segmentation\n datasets

2021/02/16 by Ekaterina Redekop, Redekop, Ekaterina, Alexey Chernyavskiy +1
Computer Science · #Adversarial Robustness in Machine Learning #Artificial Intelligence (cs.AI) #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Image and Object Detection Techniques #Machine Learning and Data Classification

paper · pdf · doi:10.48550/arxiv.2102.08021

openalex publication_date 2021/02/16 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28

Abstract

The success of modern deep learning algorithms for image segmentation heavily\ndepends on the availability of large datasets with clean pixel-level\nannotations (masks), where the objects of interest are accurately delineated.\nLack of time and expertise during data annotation leads to incorrect boundaries\nand label noise. It is known that deep convolutional neural networks (DCNNs)\ncan memorize even completely random labels, resulting in poor accuracy. We\npropose a framework to train binary segmentation DCNNs using sets of unreliable\npixel-level annotations. Erroneously labeled pixels are identified based on the\nestimated aleatoric uncertainty of the segmentation and are relabeled to the\ntrue value.\n

Related

• Automated Detection of Label Errors in Semantic Segmentation Datasets via Deep Learning and Uncertainty Quantification
• Weakly-Supervised Semantic Segmentation by Learning Label Uncertainty
• Uncertainty Based Detection and Relabeling of Noisy Image Labels
• Learning to Segment from Noisy Annotations: A Spatial Correction Approach
• Learning Confident Classifiers in the Presence of Label Noise