{ "id": 294898, "title": "Accept uncertainty of labels: Bayesian inference", "url": "/competitions/sartorius-cell-instance-segmentation/discussion/294898", "author_name": "John Doe", "post_date": "2021-12-13T12:27:47.090000", "votes": 2, "comment_count": 0, "views": 0, "content": "

The paper \"UNCERTAINTY-BASED METHOD FOR IMPROVING POORLY LABELED SEGMENTATION DATASETS\" approaches noisy mask relabeling problem using Bayesian inference.
\nBe careful. This paper is not for inference but for relabeling.
\nBut if we could obtain relabeled dataset for inference by this method, we may improve the accuracy.

\n

Key passage

\n

I picked up several key passages from the paper and emphasized the key phrases in bold.

\n


In image segmentation, noisy labels refer to inaccuracies or ambiguities of mask boundaries. The major sources of label noise include inter-observer variability due to human subjectivity, random mistakes made by human annotators, and errors in computer-generated labels.

\n


The authors used uncertainty estimation for iterative detection and filtering of noisy labels, which showed promising performance for the image classification task. In this work we extend this approach to the task of binary image segmentation by using pixel-level uncertainty estimation to detect and relabel noisy ground truth masks for dermatoscopy and liver CT data.

\n


It is assumed that the segmentation masks given in the training set are inherently noisy, otherwise clean ground truth masks should be artificially deteriorated.

\n


Automatic relabeling improves segmentation quality without overfitting to data inaccuracies. The proposed algorithm can be used to generate cleaner datasets for training other deep learning algorithms without having to consider the impact of noisy labels.

", "messages": [ { "id": 1616448, "postDate": "2021-12-13T12:27:47.090Z", "content": "

The paper \"UNCERTAINTY-BASED METHOD FOR IMPROVING POORLY LABELED SEGMENTATION DATASETS\" approaches noisy mask relabeling problem using Bayesian inference.
\nBe careful. This paper is not for inference but for relabeling.
\nBut if we could obtain relabeled dataset for inference by this method, we may improve the accuracy.

\n

Key passage

\n

I picked up several key passages from the paper and emphasized the key phrases in bold.

\n


In image segmentation, noisy labels refer to inaccuracies or ambiguities of mask boundaries. The major sources of label noise include inter-observer variability due to human subjectivity, random mistakes made by human annotators, and errors in computer-generated labels.

\n


The authors used uncertainty estimation for iterative detection and filtering of noisy labels, which showed promising performance for the image classification task. In this work we extend this approach to the task of binary image segmentation by using pixel-level uncertainty estimation to detect and relabel noisy ground truth masks for dermatoscopy and liver CT data.

\n


It is assumed that the segmentation masks given in the training set are inherently noisy, otherwise clean ground truth masks should be artificially deteriorated.

\n


Automatic relabeling improves segmentation quality without overfitting to data inaccuracies. The proposed algorithm can be used to generate cleaner datasets for training other deep learning algorithms without having to consider the impact of noisy labels.

", "rawMarkdown": "The paper \"[UNCERTAINTY-BASED METHOD FOR IMPROVING POORLY LABELED SEGMENTATION DATASETS](https://arxiv.org/pdf/2102.08021.pdf)\" approaches noisy mask relabeling problem using [**Bayesian inference**](https://en.wikipedia.org/wiki/Bayesian_inference).\nBe careful. This paper is not for inference but for **relabeling**.\nBut if we could obtain relabeled dataset for inference by this method, we may improve the accuracy.\n\n### Key passage\nI picked up several key passages from the paper and emphasized the key phrases in bold.\n\n\n
*In image segmentation, noisy labels refer to inaccuracies or ambiguities of mask boundaries. The major sources of label noise include inter-observer variability due to human subjectivity, **random** mistakes made by human annotators, and errors in computer-generated labels.*
\n\n\n
*The authors used **uncertainty estimation** for iterative detection and filtering of noisy labels, which showed promising performance for the image classification task. In this work we extend this approach to the task of binary image segmentation by using pixel-level uncertainty estimation to detect and relabel noisy ground truth masks for dermatoscopy and liver CT data.*
\n\n\n
*It is assumed that the segmentation masks given in the training set are **inherently noisy**, otherwise clean ground truth masks should be artificially deteriorated.*
\n\n\n
*Automatic relabeling improves segmentation quality without overfitting to data inaccuracies. The proposed algorithm can be used to generate cleaner datasets for training other deep learning algorithms without having to consider the impact of noisy labels.*
", "votes": 2 } ], "comments": [], "raw_markdown_by_id": { "1616448": "The paper \"[UNCERTAINTY-BASED METHOD FOR IMPROVING POORLY LABELED SEGMENTATION DATASETS](https://arxiv.org/pdf/2102.08021.pdf)\" approaches noisy mask relabeling problem using [**Bayesian inference**](https://en.wikipedia.org/wiki/Bayesian_inference).\nBe careful. This paper is not for inference but for **relabeling**.\nBut if we could obtain relabeled dataset for inference by this method, we may improve the accuracy.\n\n### Key passage\nI picked up several key passages from the paper and emphasized the key phrases in bold.\n\n\n
*In image segmentation, noisy labels refer to inaccuracies or ambiguities of mask boundaries. The major sources of label noise include inter-observer variability due to human subjectivity, **random** mistakes made by human annotators, and errors in computer-generated labels.*
\n\n\n
*The authors used **uncertainty estimation** for iterative detection and filtering of noisy labels, which showed promising performance for the image classification task. In this work we extend this approach to the task of binary image segmentation by using pixel-level uncertainty estimation to detect and relabel noisy ground truth masks for dermatoscopy and liver CT data.*
\n\n\n
*It is assumed that the segmentation masks given in the training set are **inherently noisy**, otherwise clean ground truth masks should be artificially deteriorated.*
\n\n\n
*Automatic relabeling improves segmentation quality without overfitting to data inaccuracies. The proposed algorithm can be used to generate cleaner datasets for training other deep learning algorithms without having to consider the impact of noisy labels.*
" } }