CU3D

From Computational Cognitive Neuroscience Wiki
Jump to: navigation, search


Overview

The CU3D-100 is a high-quality visual object recognition dataset suitable for studying properties of object recognition that require well-controlled images (e.g., object invariance, feature complexity). Note that this is a significant departure from datasets that contain natural images (e.g., Caltech101/256, PASCAL) which, while ecologically valid and suitable for computer vision applications, are vastly underconstrained for studying the hard problem of object recognition that the brain solves (Pinto, Cox, DiCarlo, 2008, PLoS Computational Biology). The current release of the dataset contains 100 classes of objects with an average of 9.42 exemplars per class. The images are rendered from three-dimensional models of real-world objects that were obtained from the internet and are normalized for differences in position, orientation, scale, etc. Currently, we are providing releases of the images only via this website. Any person interested in obtaining the models and renderer should contact dean [dot] wyatte [at] colorado [dot] edu.

If you use this dataset, please cite:

  • O'Reilly, R.C., Wyatte, D., Herd, S., Mingus, B., & Jilk, D. (2013). Recurrent processing during object recognition. Frontiers in Psychology, 4(124), 1-14.

Gallery of examples


Nine of the 100 object categories.

Nine exemplars from the fish category.


Nine depth, tilt, and lighting variations for one exemplar.

Nine affine transformations for a single view.

One example exemplar from each of the 100 categories

News and updates

  • 5/27/11: CU3D-100 website gets a facelift

Documentation and useful links

Download the CU3D-100


Main Release

  • Download: CU3D_100_renders_lr20_u30_nb.tar.gz (354 MB)
  • Description: 320x320 color images. 100 object classes, 20 images per exemplar. Rendered with 40° depth rotation about y-axis (plus horizontal flip), 20° tilt rotation about x-axis, 80° overhead lighting rotation. 18840 total images. The main release is not rendered with affine transformations (translation, scale, rotation). If you want to test invariance to these transformations, you will need to implement them yourself (easy using Matlab) or download the xform1 release below.
Fish 001 00001.png Fish 001 00002.png

xform0

  • Download: CU3D_100_renders_lr20_u30_nb_xform0.tar.gz (102 MB)
  • Description: 144x144 grayscale images. 100 object classes, 20 images per exemplar. Rendered with 40° depth rotation about y-axis (plus horizontal flip), 20° tilt rotation about x-axis, 80° overhead lighting rotation. 18840 total images.
Fish 001 00001 xf00 xform0.png Fish 001 00002 xf00 xform0.png

xform1

  • Download: CU3D_100_renders_lr20_u30_nb_xform1.tar.gz (123 MB)
  • Description: 144x144 grayscale images. 100 object classes, 20 images per exemplar. Rendered with 40° depth rotation about y-axis (plus horizontal flip), 20° tilt rotation about x-axis, 80° overhead lighting rotation AND 30% planar translation, 25% scaling, 14° planar rotation. 18840 total images.
Fish 001 00001 xf00 xform1.png Fish 001 00002 xf00 xform1.png

xform50

  • Download: CU3D_100_renders_lr20_u30_nb_xform50.tar (6.9 GB)
  • Description: 144x144 grayscale images. 100 object classes, 20 images per exemplar, 50 transformations per image. Rendered with 40° depth rotation about y-axis (plus horizontal flip), 20° tilt rotation about x-axis, 80° overhead lighting rotation. 944,000 total images.

3D Models


Sample OID file

  • Download: CU3D_100_renders_lr20_u30_nb.dat (931 KB)
  • Description: Sample tab-delimited file that contains the correct category label for each image in the main release (above).