Unsupervised Pre-Training of Image Features on Non-Curated Data

Caron, Mathilde; Bojanowski, Piotr; Mairal, Julien; Joulin, Armand

doi:10.1109/iccv.2019.00305

Cited by 397 publications

(526 citation statements)

References 28 publications

Supporting

Mentioning

523

Contrasting

Unclassified

Order By: Relevance

“…However, when we do not have access to such ground-truth labels, we need to define a prior to obtain an estimate of which samples are likely to belong together, and which are not. End-to-end learning approaches have utilized the architecture of CNNs as a prior [53,6,51,16,4,5], or enforced consistency between images and their augmentations [23,20] to disentangle the clusters. In both cases, the cluster learning is known to be sensitive to the network initialization.…”

Section: Representation Learning For Semantic Clusteringmentioning

confidence: 99%

“…DEC [51], DAC [6], DeepCluster [4], DeeperCluster [5], or others [1,16,53]) leverage the architecture of CNNs as a prior to cluster images. Starting from the initial feature representations, the clusters are iteratively refined by deriving the supervisory signal from the most confident samples [6,51], or through cluster re-assignments calculated offline [4,5]. A second group of methods (e.g.…”

Section: Introduction and Prior Workmentioning

confidence: 99%

“…IIC [23], IMSAT [20]) propose to learn a clustering function by maximizing the mutual information between an image and its augmentations. In general, methods that rely on the initial feature representations of the network are sensitive to initialization [6,51,4,5,21,16,53], or prone to degenerate solutions [4,5], thus requiring special mechanisms (e.g. pretraining, cluster reassignment and feature cleaning) to avoid those situations.…”

Section: Introduction and Prior Workmentioning

confidence: 99%

See 2 more Smart Citations

SCAN: Learning to Classify Images Without Labels

Gansbeke

Vandenhende

Georgoulis

et al. 2020

Computer Vision – ECCV 2020

270

293

View full text Add to dashboard Cite

Can we automatically group images into semantically meaningful clusters when ground-truth annotations are absent? The task of unsupervised image classification remains an important, and open challenge in computer vision. Several recent approaches have tried to tackle this problem in an end-to-end fashion. In this paper, we deviate from recent works, and advocate a two-step approach where feature learning and clustering are decoupled. First, a self-supervised task from representation learning is employed to obtain semantically meaningful features. Second, we use the obtained features as a prior in a learnable clustering approach. In doing so, we remove the ability for cluster learning to depend on low-level features, which is present in current end-to-end learning approaches. Experimental evaluation shows that we outperform stateof-the-art methods by large margins, in particular +26.6% on CIFAR10, +25.0% on CIFAR100-20 and +21.3% on STL10 in terms of classification accuracy. Furthermore, our method is the first to perform well on a largescale dataset for image classification. In particular, we obtain promising results on ImageNet, and outperform several semi-supervised learning methods in the low-data regime without the use of any ground-truth annotations. The code is available at www.github.com/wvangansbeke/ Unsupervised-Classification.git.

show abstract

Section: Representation Learning For Semantic Clusteringmentioning

confidence: 99%

Section: Introduction and Prior Workmentioning

confidence: 99%

Section: Introduction and Prior Workmentioning

confidence: 99%

See 1 more Smart Citation

SCAN: Learning to Classify Images Without Labels

Gansbeke

Vandenhende

Georgoulis

et al. 2020

Computer Vision – ECCV 2020

270

293

View full text Add to dashboard Cite

show abstract

“…As discussed in Section 2.2, standard deep learning models cannot be directly used for unsupervised tasks, where we do not have the labels of images. Self-supervised learning of image features has been recently proposed in deep learning community (Yang et al, 2016;Gidaris et al, 2018;Caron et al, 2018Caron et al, , 2019. 8 The intuition behind self-supervised learning is that we can perform data augmentation and create pseudo-categories (Gidaris et al, 2018 We then apply a predetermined clustering algorithm such as k-means on the initial low-dimensional vectors to group them into K clusters.…”

Section: Self-supervised Learning Of Image Representationsmentioning

confidence: 99%

Image Clustering: An Unsupervised Approach to Categorize Visual Data in Social Science Research

Zhang¹,

Peng²

2021

Preprint

View full text Add to dashboard Cite

Automated image analysis has received increasing attention in social scientific research, yet existing scholarship has focused on the application of supervised machine learning to classify images into predefined categories. This study focuses on the task of unsupervised image clustering, which automatically finds categories from image data. First, we review the steps to perform image clustering, and then we focus on the key challenge of performing unsupervised image clustering---finding low-dimensional representations of images. We present several methods of extracting low-dimensional representations of images, including the traditional bag-of-visual-words model, self-supervised learning, and transfer learning. We compare these methods using two datasets containing images related to protests in China (from Sina Weibo, Chinese Twitter) and to climate change(from Instagram). Results show that transfer learning significantly outperforms other methods. The dataset used in the pretrained model critically determines what categories algorithms can discover.

show abstract

“…This learning process uses a computer to classify data randomly. This process is usually known as the concept of unsupervised learning [29] [30]. The input is data or objects and clusters.…”

Section: K-means Clusteringmentioning

confidence: 99%

K-Means Clustering for Egg Embryo's Detection Based-on Statistical Feature Extraction Approach of Candling Eggs Image

Saifullah

2020

Sinergi

View full text Add to dashboard Cite

This research discusses the detection of embryonic eggs using the k-means clustering method based on statistical feature extraction. The processes that occur in detection are image acquisition, image enhancement, feature extraction, and identification/detection. The data used consisted of 200 egg image data, consisting of 100 test data and 100 new test data. The acquisition process uses a smartphone camera by capturing candled egg objects. The results of image acquisition become a reference in the process of image enhancement and feature extraction using Statistical Feature Extraction. The statistical feature extraction applied is the Gray Level Co-occurrence Matrix (GLCM) method, which consists of 6 features, namely Energy, Contrast, Entropy, Variance, Correlation, and Homogeneity. The results of feature extraction (6 features) are grouped by the K-means Clustering method. The clustering process uses Euclidean distance calculations to determine the proximity of features. The results of grouping and testing give the best average results with an accuracy of ≈ 74% from several test samples.

show abstract

Unsupervised Pre-Training of Image Features on Non-Curated Data

Cited by 397 publications

References 28 publications

SCAN: Learning to Classify Images Without Labels

SCAN: Learning to Classify Images Without Labels

Image Clustering: An Unsupervised Approach to Categorize Visual Data in Social Science Research

K-Means Clustering for Egg Embryo's Detection Based-on Statistical Feature Extraction Approach of Candling Eggs Image

Contact Info

Product

Resources

About