COIL-RWTH

This page contains some data sets for evaluation of classifiers that detect objects in images. More details may also be found in the following publication:

D. Keysers, M. Motter, T. Deselaers, and H. Ney. Training and Recognition of Complex Scenes using a Holistic Statistical Model. In DAGM 2003, Pattern Recognition, 25th DAGM Symposium, Magdeburg, Germany, Volume LNCS 2781 of Lecture Notes in Computer Science, pages 52-59, September 2003. ©Springer-Verlag. (.pdf)

The dataset is based on the COIL-20 dataset from the Columbia Object Image Library as described in the technical report available from the page above: "Columbia Object Image Library (COIL-20)," S. A. Nene, S. K. Nayar and H. Murase, Technical Report CUCS-005-96, February 1996.

We use the train objects from the COIL-20 "processed" corpus with odd 3D-angles (i.e. odd image numbers from 0-71) and the test objects are from the COIL-20 "unprocessed" corpus, even 3D-angles (i.e. even image numbers from 0-71). The unprocessed corpus has a different image resolution and lighting conditions are also different, which makes the task more realistic. (Other publications use a split by angle of the "processed" corpus.) Unfortunately, the unprocessed corpus only contains 5 of the 20 images. This fact should not be used by the classifier.

These are the data sets:

COIL-RWTH-1.tar.gz (22M) (train and test data with black background)
COIL-RWTH-2.tar.gz (118M) (train and test data with heterogeneous background)

The COIL-RWTH-1 corpus contains objects placed on a homogeneous black background, whereas the COIL-RWTH-2 corpus contains the objects in front of inhomogeneous real-world background images that were kept separate for training and test images and vary in resolution. The two training and test sets are based on the COIL-20 sets as described above. The training images are of size 192x192 and the size of the test images is 448x336. In all sets, we applied the following uniformly distributed random transformations to the object images: translation, 360 degree 2D-rotation, and 60%-100% scaling with fixed aspect ratio.

Here is a summary of results:

train: COIL-20 processed, odd 3D-angles/image numbers
test: COIL-20 unprocessed, even 3D-angles/image numbers
result: 0% error rate
train: COIL-RWTH-1-TRAIN
test: COIL-20 unprocessed, even 3D-angles/image numbers
result: 7.8% error rate
train: COIL-RWTH-1-TRAIN
test: COIL-20 unprocessed, even 3D-angles/image numbers
specials: 2-D rotation angles known in training
result: 4.4% error rate
train: COIL-20 processed, odd 3D-angles/image numbers
test: COIL-RWTH-1-TEST
specials: 2-D rotation angles known
result: 1.1% error rate
train: COIL-RWTH-1-TRAIN
test: COIL-RWTH-1-TEST
specials: 2-D rotation angles known
result: 7.8% error rate
train: COIL-RWTH-2-TRAIN
test: COIL-RWTH-2-TEST
specials: 2-D rotation angles known
result: 92.8% error rate

If you use this data please reference the above publication and the COIL-20 dataset. If you obtain interesting results or are interested in additional results not published here, please contact us, e.g. Daniel Keysers.

Daniel Keysers

Last modified: Thu Jun 3 22:02:39 CEST 2004