Computer Vision Datasets

Question

I am testing computer vision algorithms to categorize images. I would like to find a dataset with several categories of objects, for example. cats and dogs. This dataset must have all the variability in each class due to the intrinsic variability of the class. That is, I do not want to worry about photographs taken from different points of view or under different lighting conditions, etc. Almost all variability in a category should be due to the internal variability of this category, for example. the cat category will contain many different images, because the cats really look different from each other, and not because the images were created in different conditions.

Preferably, the objects are “cut out” (against a uniform background). The size of the data set is not important. Synthetic images (possibly created using 3D graphics software) are also approved. Images should be tagged with their category.

Does anyone know about such a dataset?

Answer 1

Computer Vision Online seems to have a very complete list of CV datasets and image processing: http://computervisiononline.com/datasets

Answer 2

I tried using SVM in this library a few years ago:

http://www.cs.columbia.edu/CAVE/software/softlib/coil-100.php

Answer 3

Try the caltech 256 dataset: http://www.vision.caltech.edu/Image_Datasets/Caltech256/

This will give you a good basis for comparison with other algorithms used in the community.

Answer 4

I have listed some of the image datasets that we worked on during my PhD, but you should really find a lot of them on the Internet. From what you described, you are looking for an object recognition task or segmentation dataset using groundtruth.

You might be interested in ALOI Dataset : “ALOI is a collection of color images from thousands of objects recorded for scientific purposes. To capture sensory variation in recording objects, we systematically changed the viewing angle, the angle of illumination and the color of the backlight for each object and additionally captured wide - main stereo images. We recorded over a hundred images of each object, resulting in a collection of 110,250 images. "

A traffic sign recognition identifier may also be of interest to you. IIRC there is also a terrestrial masking of truth for road signs.

In both cases, you should be able to replace the background with what you want (if you want to make the task more complicated).

Good luck in the recognition problem (if it is still relevant).

Answer 5

I think the CIFAR dataset can help you. It says it includes tagged subsets of 80 million tiny sets of images. Loading a dataset seems easy in Python, Matlab, or C (binary version) using the provided methods.

Answer 6

That is, I do not want to worry about photographs taken from different points of view or under different lighting conditions, etc. Almost all variability in a category should be due to the internal variability of this category, for example. The category of cats contain many different images, because cats really look different from each other, not because the images were created under different conditions.

This is not very realistic. In fact, you will have a system that does not work at all in practice - or only in very limited environments.

One such dataset that I know (I don't need such datasets) is the UIUC car dataset . This database contains images of side views of cars for use in evaluating object detection algorithms. (single-scale + multi-scale)

Answer 7

Try PASCAL, ImageNet, and SUN. Caltech and MSRC are old and off.

You might also want to read this article: An unbiased look at dataset bias - Torralba (MIT) and Efros (CMU): people.csail.mit.edu/torralba/publications/datasets_cvpr11.pdf