Image Recognition Benchmarks: MNIST, CIFAR, and Caltech Series

This GitHub repository curated by chatflip aggregates six standard computer vision benchmarks including Caltech101/256, CIFAR-10/100, and MNIST/FashionMNIST as of March 2023. It serves as a centralized collection of image classification data ranging from handwritten digits to complex object categories.

Use Cases

• Benchmarking image classification models using CIFAR-100 object labels
• Training generative adversarial networks (GANs) on MNIST digit structures
• Evaluating transfer learning performance using Caltech256 category labels

Strengths

• Includes 6 distinct benchmark datasets
• Covers class counts ranging from 10 to 256
• Provides both grayscale and RGB color image formats

Limitations

• Aggregated from external sources without original metadata
• Lack of documentation on specific file formats or directory structure
• No information provided regarding train/test split consistency

Provenance

Source

Caltech, University of Toronto (CIFAR), and Zalando (FashionMNIST)

Collection Method

Aggregation

Freshness

Last updated March 2023; contains historical benchmarks.