This page includes a list of curated publicly available datasets for machine learning and computer graphics research projects.
Hosting and distributing large datasets is cumbersome and expensive in storage and bandwidth costs. The torrent protocol for file transfer (while unfortunately synonymous with internet piracy and theft) actually offers a perfect solution to the problem of hosting and allowing access to extremely large datasets.
Once you have downloaded a dataset if you have the means and capacity you should seed for a reasonable amount of time to allow others to benefit from your having downloaded the data.
| Dataset | Description | Link |
|---|---|---|
| Vincent van Gogh Paintings | High resolution scans of Vincent van Gogh Paintings. | Link |
| Arizona State University Twitter Data Set | A large dataset of twitter users, their friend and follower relationships, and connectivities based on tweets. | Link |
| Mars Terrain Height Data | 46080x23040 high resolution mosaic of Martian terrain height data collected by the Mars Orbiter Laser Altimeter |
Link |
| IARPA Functional Map of the World | 2.69TB of high resolution satellite imagery covering the entire globe. |
Link |
Small scale image datasets useful for prototyping and experimentation. Datasets are small enough to fit in RAM without causing issues, even on small formfactor embedded and mobile devices, and on low power laptops. Useful for small scale image synthesis and classification.
| Dataset | Description | Link |
|---|---|---|
| MNIST | Labelled 28x28 grayscale handwritten digits in 0-9 from 10 classes. |
http://yann.lecun.com/exdb/mnist/ |
| notMNIST | Labeled 28x28 grayscale images of letters in the range A-J from 10 classes sampled from publicly available fonts. |
http://yaroslavvb.com/upload/notMNIST/ |
| Fashion MNIST | Labelled 28x28 grayscale images of clothing items. |
https://github.com/zalandoresearch/fashion-mnist |
| CIFAR-10 and CIFAR-100 | 60,000 labelled 32x32 RGB images from 10 or 100 classes, respectively. |
https://www.cs.toronto.edu/~kriz/cifar.html |
Larger scale image datasets which are too big to fit in RAM. Loading data efficiently during training and inference will be vital when working with these. Useful for image synthesis, classification, and segmentation.
| Dataset | Description | Link |
|---|---|---|
| SVHN | ~600,000 labelled 32x32 RGB images of Google Street View House Numbers. |
http://ufldl.stanford.edu/housenumbers/ |
| ImageNet | ~14 million labelled RGB images with an average resolution of 469x387 from 1000 classes. Human readable class names. |
http://www.image-net.org/ |
| Cityscapes | High resolution RGB images of drivers point of view within 50 different cities. 5000 images with fine segmentation masks, 20,000 with coarse segmentation masks. |
https://www.cityscapes-dataset.com/ |
| CelebA | ~200,000 annotated 178x218 RGB images of celebrity faces, cropped and aligned consistently. Licensing and usage rights on this data is ambiguous or non-existent. |
http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html |
| FFHQ | 70,000 RGB images of human faces at 1024x1024 resolution. All images are under a permissive license. This dataset is the logical successor to CelebA due to quality and licensing. |
https://github.com/NVlabs/ffhq-dataset |
Large scale dataset of 3D models of various common object types. Useful for mesh synthesis, classification, and segmentation.
| Dataset | Description | Link |
|---|---|---|
| ShapeNetCore | 51,300 3D meshes labelled by 55 object categories. |
https://www.shapenet.org/ |
| ShapeNetSem | 12,000 3D meshes, sampled from ShapeNetCore, semantically labelled per triangle face into 270 component categories. |
https://www.shapenet.org/ |
Small scale but high quality (polycount) datasets of 3D models useful for testing rendering applications. The classic Utah Teapot, Stanford Bunny, Stanford Dragon, Buddha, and Armadillo meshes can be found in these repositories.
| Dataset | Description | Link |
|---|---|---|
| The Stanford 3D Scanning Repository | High resolution 3D meshes created via laser scanning of figurines. Meshes are raw from scan and may require clean up! Rendering a mesh with holes using a glass shader will allow light to leak out of the mesh without scattering properly. | http://graphics.stanford.edu/data/3Dscanrep/ |
| McGuire Computer Graphics Archive | High quality curated models and scenes with texture and bump maps. | http://casual-effects.com/data/ |
| Keenan's 3D Model Repository (Carnegie Mellon University) | Nice toy meshes and the occasional high resolution laser scan. | https://www.cs.cmu.edu/~kmcrane/Projects/ModelRepository/ |
| Large Geometric Models Archive (Georgia Tech) | More high resolution scans. | https://www.cc.gatech.edu/projects/large_models/ |