Uncertainty-based method for improving poorly labeled segmentation datasets

TL;DR

This paper introduces an uncertainty-based framework to improve the training of segmentation neural networks on datasets with noisy labels by identifying and correcting mislabeled pixels using estimated aleatoric uncertainty.

Contribution

It presents a novel method that leverages aleatoric uncertainty to detect and relabel incorrect annotations, enhancing segmentation accuracy on poorly labeled datasets.

Findings

Improved segmentation accuracy on noisy datasets

Effective identification of mislabeled pixels

Enhanced robustness of segmentation models

Abstract

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