Computer Vision Group Department of Computer Science University of Illinois at Urbana-Champaign

Overview

The CORE dataset is intended to help learn more detailed models and for exploring cross-category generalization in object recognition. It has already been used for exploring problems such as:

Describing familiar and unfamiliar objects

Localizing parts for pose, viewpoint, and object parsing

News

11/1/2010 - Included evaluation code and precomputed detections.

10/12/2010 - Updated Dataset README

The README is also included with the annotations.

6/11/2010 - The website is now up!

Everything you need to get started with the dataset is here; learning and inference tools will be added soon.

Be sure to visit our workshop talk at ACVHL2010 and poster at CVPR2010 this week.

Publications

[1] “Attribute-Centric Recognition for Cross-category Generalization” [pdf]

Ali Farhadi, Ian Endres, and Derek Hoiem

CVPR 2010.

[2] “The Benefits and Challenges of Collecting Richer Object Annotations” [pdf]

Ian Endres, Ali Farhadi, Derek Hoiem, and David Forsyth

ACVHL 2010 (in conjunction with CVPR).

[3] “Describing Objects by their Attributes” [pdf][related website]

Ali Farhadi, Ian Endres, Derek Hoiem, and David A. Forsyth

CVPR 2009.

Browse

Downloads

Annotated Dataset CORE_v1_data.tar.gz (522MB)

The base dataset, includes the images and annotations.

The LabelMe toolbox provides many helpful functions for manipulating this dataset: LabelMe Toolbox

Detection Evaluation CORE_v1_eval.tar.gz (7.2KB)

Code used in [1] to evaluate part and object predictions on the CORE dataset.

Precomputed Bounding Boxes CORE_v1_detections.tar.gz (3.4GB)

Top 1000 bounding boxes for each image from category and part models trained with Felzenszwalb et al.

Dataset Toolbox (Coming Soon) CORE_v1_tools.tar.gz

A set of Matlab code to supplement the LabelMe toolbox. This includes code to retrieve, aggregate, and display data from the images. See documentation included in the package.

Object/Part Detection Models (Coming Soon) CORE_v1_det_models.tar.gz

These are the models we used for localizing parts and objects in [1]

These require the detection code from Felzenszwalb et al. (use version 3.x): Detection

Training, Inference, and Evaluation (Coming Soon) CORE_v1_full_model.tar.gz

This code in conjunction with the Object/Part Models can be used to reproduce the results found in [1]

Funding

This dataset was supported in part by a Google Research Award, the National Science Foundation under IIS (0904209), an NSF CAREER Award (1053768), an ONR MURI Award (N000141010934), and gifts from Microsoft and Intel.

Last Modified: