80 Million Tiny Images: A Large Data Set for Nonparametric Object and Scene Recognition

By Antonio Torralba, Rob Fergus, William T. Freeman. 2008

779 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition, Media Technology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Media Technology

Citations in the corpus

Towards neural networks that provably know when they don't know (2019, methodology): “However, in this work we follow the approach of Hendrycks et al. (2019) where they used the 80 million tiny image dataset (Torralba et al., 2008) as a proxy of all possible images.”
Class-Incremental Learning: Survey and Performance Evaluation on Image Classification (2020, methodology): “Due to the withdrawal of Tiny Images [127] (the auxiliary dataset used in the original implementation), we use a 32 ˆ 32 resized Imagenet as auxiliary dataset and we set its batch size to 128.”
On the genealogy of machine learning datasets: A critical history of ImageNet (2021, methodology): “Previous datasets, such as TinyImages (Torralba et al., 2008), had leveraged this same method of data collection and categorical structure.”
Energy-based Out-of-distribution Detection (2020, methodology): “For the auxiliary outlier dataset, we use 80 Million Tiny Images [39], which is a large-scale, diverse dataset scraped from the web.”
Incremental Learning with Unlabeled Data in the Wild (2019, methodology): “To simulate a large stream of unlabeled data, we take two large datasets: TinyImages [18] with 80M images and ImageNet 2011 with 14M images.”
Factoring Variations in Natural Images with Deep Gaussian Mixture Models (2014, methodology): “For our experiments we used the Berkeley Segmentation Dataset (BSDS300) [13], which is a commonly used benchmark for density modeling of image patches and the tiny images dataset [14].”
Learning to Learn with Compound HD Models (2011, methodology): “Similar to [8], in order to learn good generic low-level features, we first train a two-layer DBM in completely unsupervised fashion using 4 million tiny images5 [23].”
New trends and ideas in visual concept detection: the MIR flickr retrieval evaluation initiative (2010, background): “As outlined in for instance [30] the billions of images freely available on the internet provide us with a dense sampling of the visual world.”
Evaluation of GIST descriptors for web-scale image search (2009, methodology): “In addition to these images, we use the “tiny image” dataset of Torralba and Fergus [22].”
Nonparametric Method for Data-driven Image Captioning (2014, methodology): “The basic I M 2T EXT model uses an equally weighted average of GIST (Oliva and Torralba, 2001) and TinyImage (Torralba et al., 2008) features, which coarsely localize low-level features in scenes.”
Progress in wearable electronics/photonics—Moving toward the era of artificial intelligence and internet of things (2020, methodology): “…10E. 51 By analogizing the fundamental perception primitives between the visual and tactile domains, the STAG is assembled with a minimal sensor count as that for image processing (32 × 32 pixels) 302 of 584 piezoresistive sensor array distributed on palm for interacting with 26 different objects.”
Rethinking Semantic Segmentation: A Prototype View (2022, neutral): “Different from current approaches that are typically built upon learnable prototypes, in pre-deep era, many segmentation systems are nonparametric [32, 73, 74, 80, 102, 103].”
Human Uncertainty Makes Classification More Robust (2019, methodology): “The images are taken from TinyImages [46] and match the sub-class distributions in CIFAR10 . v6 has 200 images per class while v4 is the original class-unbalanced version (90% overlap).”
Vehicle detection in aerial imagery : A small target detection benchmark (2016, background): “[12, 13, 14, 15, 16]) – which is related to our problem but covers different needs – only a few of them specifically address object detection.”
BabyTalk: Understanding and Generating Simple Image Descriptions (2013, background): “image—this is true even for very low-resolution images [40]”
OpenOOD: Benchmarking Generalized Out-of-Distribution Detection (2022, background): “For OOD dataset, near-OOD includes CIFAR-10 [43] and TinyImageNet [23].”
Making Archetypal Analysis Practical (2009, methodology): “For both data sets, we followed the proposal in [15] and re-scaled the RGB/Gray-value images to a resolution of 32× 32 pixels.”
Semi-Supervised Hashing for Large-Scale Search (2012, methodology): “The CIFAR-10 dataset is a labeled subset of the 80-million tiny images collection [31].”