Image classification has become a popular task in computer vision. In recent years, new advances in machine learning and deep learning have led to improved the classification results. Additionally, large amount of image classification datasets has been compiled and made available. When you work with deep learning algorithms, you eventually need a lot of data to train your complex models without overfitting.
Table of Contents
Image classification has become a popular task in computer vision. In recent years, new advances in machine learning and deep learning have led to improved the classification results. Additionally, large amount of image classification datasets has been compiled and made available. When you work with deep learning algorithms, you eventually need a lot of data to train your complex models without overfitting.
It is a very tedious task to collect your own data and assign the labels to them for the image classification task. Fortunately, there are several publicly available datasets to help you get started with deep learning in computer vision tasks. For today’s blog post, we’ll look at the best, free, publicly available datasets for image classification task.
Large Datasets for Image Classification
When it comes to image classification, the bigger the dataset, the better. This is because more data allows your machine learning or deep learning model to learn more patterns and generalize better. You can also utilize the large dataset for transfer learning.
iNaturalist
contains 579,184 images for training and 95,986 images for testing. The goal of this dataset is to classify the images into 5,089 species of plants and animals. iNaturalist is about 239 GB in size.
ImageNet
is most popular and large dataset for imaging tasks. It has 1,281,167 images for training and 150,000 images for validation and testing. The goal of the dataset to classify the images into 1000 categories. The dataset size is about 150 GB.
BigEarthNet
The BigEarthNet dataset is a large-scale, publicly available dataset of satellite imaging. It contains data from two radars Sentinel-1 and Sentinel-2 from the 10 countries of Europe. The dataset has been used in a number of papers and is a great resource for anyone looking to get into satellite image classification. The dataset has 590,326 non-overlapping image patches. The total size of BigEarthNet dataset is 121 GB.
Places205
This dataset contains over 2.5 million images from 205 different scene categories, which is a great example of an image classification dataset. The dataset includes high-resolution images and labels indicating the scene classification. This dataset is ideal for the training of machine learning models to recognize different scenes.
Food-101
The goal of this dataset is to classify the images into 101 food categories. It has total of 101,000 colored images of dishes. Each food class has 750 training images and 250 testing images. The total size of dataset is 5 GB.
Small Datasets for Image Classification
The small size datasets for image classification task has some advantage over large ones. You can easily work and manipulate small dataset because it required small storage space. Another benefit of the small dataset is you can quickly train and test your machine learning algorithms. Let’s check out some of the best small datasets for image classification task.
The Street View House Numbers (SVHN) Dataset
The Street View House Numbers (SVHN) Dataset is provided the images of digits from real world. The size of training data is 385 MB which has 73,257 digits. The testing data is 264 MB which has 26,032 digits. It is categorized into 10 digits from zero to nine. It has low resolution images of 32×32 size.
CIFAR-10 and CIFAR-100
both datasets contains 50,000 images for training and 10000 testing images. The only difference between these two datasets is; CIFAR-10 has 10 classes of natural images and CIFAR-100 is divided into 100 classes. All images are colored RGB images and the resolution is 32 x 32. The dataset size is 163 MB.
Oxford 102 Flower
The Oxford 102 Flower Dataset is a great dataset for those looking to get into image classification. It consists of 102 different classes of flowers, with each class having between 40 and 258 images. This dataset is perfect for those who want to get into image classification, but don’t want to deal with a huge dataset like ImageNet. The size of the dataset is 329 MB only.
Fashion-MNIST
The Fashion-MNIST dataset is a dataset of Zalando’s product images. The dataset is divided into a set of 60,000 training examples and 10,000 test images. Each example is a 28×28 grayscale image associated with one of 10 classes. This is a small dataset of approximately 32 MB of size. The Fashion-MNIST data set is much more challenging than the more commonly used handwritten digits MNIST. It’s a good dataset to start your computer vision adventure.
PlantVillage
This is another small dataset for beginners. The dataset contains 61,486 images of plants and their backgrounds. The goal of this dataset to classify images into one of the 39 plants. The size of the dataset is 825 MB. The contributor of this dataset also provide plant dataset with augmentation.
Additional Resources
You may also check other blog posts to excel in computer vision.
