Learning from Weak and Noisy Labels for Semantic Segmentation

Lu, Zhiwu; Fu, Zhenyong; Xiang, Tao; Han, Peng; Wang, Liwei; Gao, Xin

doi:10.1109/tpami.2016.2552172

Cited by 113 publications

(57 citation statements)

References 54 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…They obtained best results using region proposal algorithms to create semantic segmentation training data directly from bounding boxes. Lu et al modelled this problem as a simultaneous learning and denoising task through a convex optimization problem [31]. Ahn and Kwak proposed combining class activation maps, random walk and a learned network that predicts if pixels belong to the same region to perform semantic segmentation from image level labels [1].…”

Section: Related Workmentioning

confidence: 99%

Guided Anisotropic Diffusion and Iterative Learning for Weakly Supervised Change Detection

Daudt

Saux²,

Boulch³

et al. 2019

2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

View full text Add to dashboard Cite

Large scale datasets created from user labels or openly available data have become crucial to provide training data for large scale learning algorithms. While these datasets are easier to acquire, the data are frequently noisy and unreliable, which is motivating research on weakly supervised learning techniques. In this paper we propose an iterative learning method that extracts the useful information from a large scale change detection dataset generated from open vector data to train a fully convolutional network which surpasses the performance obtained by naive supervised learning. We also propose the guided anisotropic diffusion algorithm, which improves semantic segmentation results using the input images as guides to perform edge preserving filtering, and is used in conjunction with the iterative training method to improve results.

show abstract

Section: Related Workmentioning

confidence: 99%

Guided Anisotropic Diffusion and Iterative Learning for Weakly Supervised Change Detection

Daudt

Saux²,

Boulch³

et al. 2019

2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

View full text Add to dashboard Cite

show abstract

“…However, this method was proposed for image clustering and does not apply to pixel-level label-ing. Other researchers have considered techniques based on non-negative matrix factorization to infer the pixellevel labels of pre-segmented regions (known as superpixels) within different images (Niu et al, 2015;Lu et al, 2017). Furthermore, Zhang and Gong (2016) proposed a non-negative matrix co-factorization based approach that jointly learns a discriminative dictionary and a linear classifier that classifies features from segmented images into different classes.…”

Section: Weakly-supervised Labelingmentioning

confidence: 99%

Structure label prediction using similarity-based retrieval and weakly supervised label mapping

2019

View full text Add to dashboard Cite

Structure label prediction using similaritybased retrieval and weakly supervised label mapping", GEOPHYSICS 2019 84:1, V67-V79. ABSTRACTRecently, there has been significant interest in various supervised machine learning techniques that can help reduce the time and effort consumed by manual interpretation workflows. However, most successful supervised machine learning algorithms require huge amounts of annotated training data. Obtaining these labels for large seismic volumes is a very timeconsuming and laborious task. We address this problem by presenting a weakly-supervised approach for predicting the labels of various seismic structures. By having an interpreter select a very small number of exemplar images for every class of subsurface structures, we use a novel similarity-based retrieval technique to extract thousands of images that contain similar subsurface structures from the seismic volume. By assuming that similar images belong to the same class, we obtain thousands of image-level labels for these images; we validate this assumption in our results section. We then introduce a novel weakly-supervised algorithm for mapping these rough image-level labels into more accurate pixel-level labels that localize the different subsurface structures within the image. This approach dramatically simplifies the process of obtaining labeled data for training supervised machine learning algorithms on seismic interpretation tasks. Using our method we generate thousands of automaticallylabeled images from the Netherlands Offshore F3 block with reasonably accurate pixel-level labels. We believe this work will allow for more advances in machine learning-enabled seismic interpretation. 0

show abstract

“…Generally speaking, there are four types of existing methods: (I) robust learning based on probabilistic graphical models where the noisy patterns are often modeled as latent variables [30,25]; (II) progressive and self-paced learning, where easy and clean examples are learned first, whereas the hard and noisy labels are progressively considered [10]; (III) loss-correction methods, where the loss function is corrected iteratively [22]; (IV) network architecture-based method, where the noisy patterns are modeled with specifically designed modules [15]. Meanwhile, there are also some efforts on designing deep robust models for specific tasks and applications: [20] proposes a method to learn from Weak and Noisy Labels for Semantic Segmentation; [32] proposes a deep robust unsupervised method for saliency detection, etc.…”

Section: Learning With Noisy Datamentioning

confidence: 99%

“…When it comes to deep learning, it is known that several kinds of factors can drive the deep learning model away from a perfect one, with the data perturbation issue as an typical example. Besides the notorious issue coming from the crowdsourcing process, deep learning is in itself known to be more vulnerable to contaminated data since the extremely high model complexity brings extra risks to overfit the noisy/contaminated data [20,10,30,32,15,25,22]. We believe that how to guarantee the robustness is one of the biggest challenges when constructing deep SVP prediction models.…”

Section: Introductionmentioning

confidence: 99%

Deep Robust Subjective Visual Property Prediction in Crowdsourcing

Xu¹,

Yang

Jiang

et al. 2019

2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

View full text Add to dashboard Cite

The problem of estimating subjective visual properties (SVP) of images (e.g., Shoes A is more comfortable than B) is gaining rising attention. Due to its highly subjective nature, different annotators often exhibit different interpretations of scales when adopting absolute value tests. Therefore, recent investigations turn to collect pairwise comparisons via crowdsourcing platforms. However, crowdsourcing data usually contains outliers. For this purpose, it is desired to develop a robust model for learning SVP from crowdsourced noisy annotations. In this paper, we construct a deep SVP prediction model which not only leads to better detection of annotation outliers but also enables learning with extremely sparse annotations. Specifically, we construct a comparison multi-graph based on the collected annotations, where different labeling results correspond to edges with different directions between two vertexes. Then, we propose a generalized deep probabilistic framework which consists of an SVP prediction module and an outlier modeling module that work collaboratively and are optimized jointly. Extensive experiments on various benchmark datasets demonstrate that our new approach guarantees promising results.

show abstract

Learning from Weak and Noisy Labels for Semantic Segmentation

Cited by 113 publications

References 54 publications

Guided Anisotropic Diffusion and Iterative Learning for Weakly Supervised Change Detection

Guided Anisotropic Diffusion and Iterative Learning for Weakly Supervised Change Detection

Structure label prediction using similarity-based retrieval and weakly supervised label mapping

Deep Robust Subjective Visual Property Prediction in Crowdsourcing

Contact Info

Product

Resources

About