论文阅读 【CVPR-2022】 Adaptive Early-Learning Correction for Segmentation from Noisy Annotations

论文阅读 【CVPR-2022】 Adaptive Early-Learning Correction for Segmentation from Noisy Annotations

自适应的早期学习校正，用于从噪声注释中进行分割

studyai.com

搜索论文: [Adaptive Early-Learning Correction for Segmentation from Noisy Annotations](http://www.studyai.com/search/whole-site/?q=Adaptive Early-Learning Correction for Segmentation from Noisy Annotations)

摘要(Abstract)

Deep learning in the presence of noisy annotations has been studied extensively in classification, but much less in segmentation tasks.

存在标注噪声的深度学习任务的研究中，分类任务进行了广泛的研究，但对分割任务的研究则少得多。

In this work, we study the learning dynamics of deep segmentation networks trained on inaccurately annotated data.

在这项工作中，我们研究了在不准确的标注数据上训练的深度分割网络的学习动态特性。

We observe a phenomenon that has been previously reported in the context of classification: the networks tend to first fit the clean pixel-level labels during an “early-learning” phase, before eventually memorizing the false annotations.

我们观察到一个以前在分类方面报道过的现象：在 "早期学习 "阶段，网络倾向于首先适应干净的像素级标签，然后再最终记忆错误的标注。

However, in contrast to classification, memorization in segmentation does not arise simultaneously for all semantic categories.

然而，与分类不同的是，分割中的记忆并不是对所有语义类别都同时出现。

Inspired by these findings, we propose a new method for segmentation from noisy annotations with two key elements.

受这些发现的启发，我们提出了一种新的方法，从有噪音的标注中进行分割，有两个关键要素。

First, we detect the beginning of the memorization phase separately for each category during training. This allows us to adaptively correct the noisy annotations in order to exploit early learning.

首先，我们在训练期间为每个类别分别检测记忆阶段的开始。这使我们能够自适应地纠正带有噪声的标注，以利于早期学习。

Second, we incorporate a regularization term that enforces consistency across scales to boost robustness against annotation noise.

第二，我们加入了一个正则化项，强制在不同的尺度上执行一致性，以提高对标注噪声的鲁棒性。

Our method outperforms standard approaches on a medicalimaging segmentation task where noises are synthesized to mimic human annotation errors.

我们的方法在医学影像分割任务上的表现优于标准方法，在该任务中，通过合成噪声用来模仿人类标注错误。

It also provides robustness to realistic noisy annotations present in weakly-supervised semantic segmentation, achieving state-of-the-art results on PASCAL VOC 2012.

它还提供了对弱监督语义分割中存在的真实噪声标注的鲁棒性，在PASCAL VOC 2012上取得了最先进的结果。