Learning with noisy labels

The CIFAR-100 dataset (Canadian Institute for Advanced Research, 100 classes) is a subset of the Tiny Images dataset and consists of 60000 32x32 color images. The 100 classes in the CIFAR-100 are grouped into 20 superclasses. There are 600 images per class. Each image comes with a "fine" label (the class to which it belongs) and a "coarse" label (the superclass to which it belongs). There are 500 training images and 100 testing images per class. The criteria for deciding whether an image belongs to a class were as follows: * The class name should be high on the list of likely answers to the question “What is in this picture?” * The image should be photo-realistic. Labelers were instructed to reject line drawings. * The image should contain only one prominent instance of the object to which the class refers. * The object may be partially occluded or seen from an unusual viewpoint as long as its identity is still clear to the labeler. Source: https://www.cs.toronto.edu/~kriz/cifar.html Image Source: https://www.cs.toronto.edu/~kriz/cifar.html

The Food-101 dataset consists of 101 food categories with 750 training and 250 test images per category, making a total of 101k images. The labels for the test images have been manually cleaned, while the training set contains some noise. Source: Combining Weakly and Webly Supervised Learning for Classifying Food Images Image Source: https://data.vision.ee.ethz.ch/cvl/datasets_extra/food-101/

VoxCeleb1 is an audio dataset containing over 100,000 utterances for 1,251 celebrities, extracted from videos uploaded to YouTube.

Clothing1M contains 1M clothing images in 14 classes. It is a dataset with noisy labels, since the data is collected from several online shopping websites and include many mislabelled samples. This dataset also contains 50k, 14k, and 10k images with clean labels for training, validation, and testing, respectively. Source: Label-Noise Robust Generative Adversarial Networks Image Source: https://openaccess.thecvf.com/content_cvpr_2015/papers/Xiao_Learning_From_Massive_2015_CVPR_paper.pdf

The WebVision dataset is designed to facilitate the research on learning visual representation from noisy web data. It is a large scale web images dataset that contains more than 2.4 million of images crawled from the Flickr website and Google Images search. The same 1,000 concepts as the ILSVRC 2012 dataset are used for querying images, such that a bunch of existing approaches can be directly investigated and compared to the models trained from the ILSVRC 2012 dataset, and also makes it possible to study the dataset bias issue in the large scale scenario. The textual information accompanied with those images (e.g., caption, user tags, or description) are also provided as additional meta information. A validation set contains 50,000 images (50 images per category) is provided to facilitate the algorithmic development.

This work presents two new benchmark datasets (CIFAR-10N, CIFAR-100N), equipping the training dataset of CIFAR-10 and CIFAR-100 with human-annotated real-world noisy labels that we collect from Amazon Mechanical Turk.

This work presents two new benchmark datasets (CIFAR-10N, CIFAR-100N), equipping the training dataset of CIFAR-10 and CIFAR-100 with human-annotated real-world noisy labels that we collect from Amazon Mechanical Turk.

Chaoyang dataset contains 1111 normal, 842 serrated, 1404 adenocarcinoma, 664 adenoma, and 705 normal, 321 serrated, 840 adenocarcinoma, 273 adenoma samples for training and testing, respectively. This noisy dataset is constructed in the real scenario. - Details: Colon slides from Chaoyang hospital, the patch size is 512 × 512. We invited 3 professional pathologists to label the patches, respectively. We took the parts of labeled patches with consensus results from 3 pathologists as the testing set. Others we used as the training set. For the samples with inconsistent labeling opinions of the three doctors in the training set (this part accounts for about 40%), we randomly selected the opinions from one of the three doctors. - The original WSIs are scanned at X20 objective magnification.

10 classes with 50, 000 training and 5, 000 testing images. Please note that, in ANIMAL10N, noisy labels were injected naturally by human mistakes, where its noise rate was estimated at 8%.

Approx. 300,000 images of galaxies labelled by shape. Labels are from www.galaxyzoo.org volunteers, and are noisy. Images are from the DECaLS telescope survey. Also includes predictions from an ensemble of EfficientNets, each using MC Dropout and a novel probabilistic loss function.

Part of the Controlled Noisy Web Labels Dataset.

Part of the Controlled Noisy Web Labels Dataset.

Part of the Controlled Noisy Web Labels Dataset.

COCO-N Medium introduces a stochastic benchmark that simulates common real-world scenarios with noticeable label inaccuracies in the COCO dataset. This benchmark combines class and spatial noises to create a challenging yet realistic evaluation framework for instance segmentation models. It mimics datasets manually annotated by crowd workers, where a moderate level of label noise is expected. By incorporating both class and spatial inaccuracies, COCO-N Medium allows researchers to assess their models' basic robustness to label noise, providing insights into performance in typical real-world applications where perfect annotations are rare. This medium-level benchmark serves as a crucial middle ground, offering a more rigorous test than minimally noisy datasets while remaining within the bounds of commonly encountered data quality issues. COCO-N Medium enables a nuanced evaluation of model performance under realistic conditions, helping identify areas for improvement in handling noisy labels and guiding the development of more robust instance segmentation algorithms.

The COCO-WAN benchmark is designed to assess the impact of weakly annotations (combined with auto-annotation tools) noise on instance segmentation models. This benchmark is built upon the COCO dataset and incorporates noise generated through weak annotations, simulating real-world scenarios where annotations might be imperfect due to semi-automated tools. It includes various levels of noise to challenge the robustness and generalization capabilities of segmentation models. Accurately labeling instance segmentation datasets is a complex and error-prone task, often leading to noisy labels. The COCO-WAN benchmark aims to provide a realistic testing ground for models to handle such noisy annotations. By utilizing foundation models and weak annotations, COCO-WAN simulates semi-automated annotation tools, helping researchers understand how well their models can perform under less-than-ideal labeling conditions. This benchmark includes multiple noise levels (easy, medium, and hard) to reflect varying degrees of annotation imperfections. Potential Use Cases of the Dataset: Model Robustness Testing: Researchers can use COCO-WAN to evaluate how different instance segmentation models cope with noisy annotations, allowing for the development of more resilient algorithms. Annotation Tool Improvement: By analyzing model performance on COCO-WAN, developers of annotation tools can identify common pitfalls and work on reducing noise in their outputs. Semi-Automated Annotation Systems: The benchmark provides insights into how models trained with semi-automated annotations perform, guiding improvements in such systems for better accuracy and efficiency in labeling tasks. The COCO-WAN benchmark offers a crucial resource for advancing the field of instance segmentation by highlighting the challenges posed by noisy labels and fostering the creation of more robust and reliable models. Model Robustness Testing: Researchers can use COCO-WAN to evaluate how different instance segmentation models cope with spatial, real noisy annotations, allowing for the development of more resilient algorithms. Semi-Automated Annotation Systems: The benchmark provides insights into how models trained with semi-automated annotations perform, guiding improvements in such systems for better accuracy and efficiency in labeling tasks. The COCO-WAN benchmark offers a crucial resource for advancing the field of instance segmentation by highlighting the challenges posed by noisy labels and fostering the creation of more robust and reliable models.