Classify handwriten digits. Some additional results are available on the original dataset page.

Classify 32x32 colour images.

Classify 32x32 colour images.

Similar to CIFAR-10 but with 96x96 images. Original dataset website.

The Street View House Numbers (SVHN) Dataset.

SVHN is a real-world image dataset for developing machine learning and object recognition algorithms with minimal requirement on data preprocessing and formatting. It can be seen as similar in flavor to MNIST(e.g., the images are of small cropped digits), but incorporates an order of magnitude more labeled data (over 600,000 digit images) and comes from a significantly harder, unsolved, real world problem (recognizing digits and numbers in natural scene images). SVHN is obtained from house numbers in Google Street View images.

1000 categories classification challenge. With tens of thousands of training, validation and testing images.

See this interesting comparative analysis.

Pascal VOC 2007 is commonly used because the test set has been realased. comp3 is the objects detection competition, using only the comp3 pascal training data.

Just like comp3 but “any training data” can be used.

Pascal VOC 2010 version of the challenge. comp3 is the objects detection competition.

Just like comp3 but “any training data” can be used.

Last Pascal VOC challenge instance (2012 version had identical data).

Project website.

Evaluated using the Caltech Pedestrians toolkit. Original dataset website.

Evaluated using the Caltech Pedestrians toolkit. Only left images used. Original dataset website.

Evaluated using the Caltech Pedestrians toolkit. Original dataset website.

Evaluated using the Caltech Pedestrians toolkit. Original dataset website.

A rich dataset to evaluate multiple computer vision tasks, including cars, pedestrian and bycicles detection.

2000 poses anotated pictures from Flickr. From a selected set of activities and with the person at the center of the pictures.

One of the oldest and classic dataset for semantic labelling. 21 different categories of surfaces are considered. Despite the innacuracies in the annotations and how unbalanced the classes are, this dataset still is commonly used as reference point. Note that here we consider the original annotations (where most results are published), not the cleaned-up version.

The results are reported per-class and per-pixel (this is sometimes called “average” and “global” result, respectively).

Original dataset website

This benchmark aggregates results from 36 methods over five datasets (MSRA10K, ECSSD, THUR15K, JuddDB, and DUTOMRON).