SOIL-47: Surrey Object Image Library


Why SOIL-47?

It is often important to understand the behaviour of recognition methods when full projective or affine effects are present, in particular when only a single view is available to learn the appearance of a 3D object. With this objective in mind, a database of coloured household objects, many of the same shape, viewed over a significant portion of the viewing sphere was collected at the University of Surrey. The images show mainly multicoloured objects, many of which consisting of planar surfaces (boxes) and with generally complex colour structure. See an example of frontal views of all the objects. The database contains 24 objects with approximately planar surfaces and 22 complex scenes.

Appearance variation in SOIL-47

3D viewpoint variation
Altogether, 21 images of each of 47 objects were acquired by a robot arm moving around the object at intervals of approximately 9 degrees spanning a range up to +-90 degrees. Appearance variations are mainly due to 3D viewpoint change and self-occlusion. For example, see all views of an object from the database. Two views of the same object which differ by 90 degrees are shown below:

Illumination intensity variation
Another set of 21x47=987 images was acquired under different geometry of light sources, to test algorithms claiming invariance to changing illumination intensity. In order to separate between the two sets, we distinguish SOIL-47A and SOIL-47B, where A and B correspond to the two different lighting settings. See an example below:

Occlusion/Distractors
A set of 22 complex scenes showing more than 1 object were also imaged from 21 viewpoints. See some examples below:

Other examples of appearance variation
Images with identical colour composition but different structure:

The Data

The grabbed images are in portable network graphics (PNG) format and their size is 576x720. Due to storage space limitations, a scaled version of the original images is available on-line with size 288x360. In addition, frontal views of the objects in the original size (576x720) were included in the on-line version of the database to allow testing of algorithms claiming scale invariance. No image processing was applied to the images. For the scaling the unix utility pnmscale was used.

SOIL-47A SOIL-47B Multi-SOIL47 3.5 Mb
Frontal view 7.5 Mb Frontal view 7.5 Mb Frontal view 3.5 Mb
View 1 7.5 Mb View 1 7.5 Mb View 1 3.5 Mb
View 2 7.5 Mb View 2 7.5 Mb View 2 3.5 Mb
View 3 7.5 Mb View 3 7.5 Mb View 3 3.5 Mb
View 4 7.5 Mb View 4 7.5 Mb View 4 3.5 Mb
View 5 7.5 Mb View 5 7.5 Mb View 5 3.5 Mb
View 6 7.5 Mb View 6 7.5 Mb View 6 3.5 Mb
View 7 7.5 Mb View 7 7.5 Mb View 7 3.5 Mb
View 8 7.5 Mb View 8 7.5 Mb View 8 3.5 Mb
View 9 7.5 Mb View 9 7.5 Mb View 9 3.5 Mb
View 10 7.5 Mb View 10 7.5 Mb View 10 3.5 Mb
View 11 7.5 Mb View 11 7.5 Mb View 11 3.5 Mb
View 12 7.5 Mb View 12 7.5 Mb View 12 3.5 Mb
View 13 7.5 Mb View 13 7.5 Mb View 13 3.5 Mb
View 14 7.5 Mb View 14 7.5 Mb View 14 3.5 Mb
View 15 7.5 Mb View 15 7.5 Mb View 15 3.5 Mb
View 16 7.5 Mb View 16 7.5 Mb View 16 3.5 Mb
View 17 7.5 Mb View 17 7.5 Mb View 17 3.5 Mb
View 18 7.5 Mb View 18 7.5 Mb View 18 3.5 Mb
View 19 7.5 Mb View 19 7.5 Mb View 19 3.5 Mb
View 20 7.5 Mb View 20 7.5 Mb View 20 3.5 Mb

SOIL-47A SOIL-47B
Large frontal view 33 Mb Large frontal view 33 Mb

The full-size image set can also be downloaded (~1.8 GByte).

Related publications

Evaluating colour object recognition algorithms using the SOIL-47 database
Dimitri Koubaroulis, Jiri (George) Matas and Josef Kittler
ACCV, 2002 (to appear)
ps.gz pdf bibtex conference home
(links under construction)

Region-based representation for object recognition by relaxation labelling
A. Ahmadyfard and Josef Kittler
Advances in Pattern Recognition, 2000, , p. 297-304
ps.gz pdf bibtex conference home
(links under construction)

Acknowledgements

The SOIL-47 images were captured by Jan Burianek and Alireza Ahmadyfard at the CVSSP lab directed by Josef Kittler. The database was made public by Dimitri Koubaroulis.

Use policy

Premission is hereby granted to use the SOIL-47 image database for academic purposes only, provided that the library is refrenced in publications related to its use as follows:
J. Burianek, A. Ahmadyfard, J. Kittler: SOIL-47, The Surrey Object Image Library, Centre for Vision, Speach and Signal processing, Univerisity of Surrey, http://www.ee.surrey.ac.uk/Research/VSSP/demos/colour/soil47/

Other image sets

  • COIL-100, is a database of colour images of 100 objects collected at Columbia University. Most experiments using this database assume multi-view object recognition i.e. more than one images available to learn object appearance. The methods tested often compute shape and/or grey-scale object descriptors. In addition, many objects in the database are single-coloured and with the same or very similar colour. This poses difficulties to testing recognition algorithms that are designed to recognise multi-coloured objects. Finally, very few objects in COIL-100 consist of planar surfaces which is a requirement of some algorithms.
  • SFU data set, collected at Simon Fraser University. This dataset was compiled specifically for testing colour constancy algorithms.

Last modified: 29 Oct 2005 by J. Matas
Created: 4 Jun 2001 by Dimitrios Koubaroulis
Last modified: Tue Oct 3 18:13:27 BST 2006