FBK home > INFORMATION TECHNOLOGY > Technologies of Vision > Databases > MLDB 2013

MLDB 2013 - small object detection/recognition dataset

about MLDB 2013

* MLDB2013 is a small database of real -world images to test object detection and recognition algorithms, e.g. SIFT or MEMORI, and active contour based segmentation algorithms.

* MLDB2013 consists of a set of 12 images of objects, said references, and of a set of 12 images, said tests, where the object have to searched for. Each object is modeled by a single view (see Figure 1). This view is represented by a color image and by a binary mask, whose black pixels specify the portion of the image actually occupied by this view, while the white pixels correspond to the background (see Figure 2).

#e6f3ff;


Figure 1: The reference images of the 12 objects of MLDB2013. Figure 2: The 10th object in MLDB2013 along with its mask, that specifies the image part actually belonging to the reference object. Figure 3: The test images of MLDB2013
From Left to right - Figure 1: The reference images of the 12 objects of MLDB2013. Figure 2: The 10th object in MLDB2013 along with its mask, that specifies the image part actually belonging to the reference object. Figure 3: The test images of MLDB2013.


For any i = 1,..., 12, the i-th test image displays the i-th reference. The test images are shown in Figure 3. The position of the reference into the correspondent test image is specified by a binary mask, where - as for the references - the black pixels belong to the view portrayed in the test image, while the white ones belong to the background.

The reference objects differ in scale, in-plane orientation and pose from their occurrences in the test images. Slight color changes due to the automatic color bal ancing of the camera also occur.

image format

* The images have been captured by a camera SONY DCR -HC17E.
The reference and the test color images are in PPM format. The binary masks are saved in PBM format. Color and binary images are denoted as img%2d.ppm and img%2d.pbm respectively. %2d is an integer codified by two digits and it varies from 01 to 12. The size of the reference images varies from 291x103 (minimum) to 857 x 1142 (maximum). The minimum and maximum areas of the objects are respectively 22 288 and 393 915 pixels. The size of the test images varies from 288 x 384 (minimum) to 857 x 1142 (maximum). Memory: 5.7 Mb (references) + 11 Mb (tests).

download MLDB2013

The complete database, with ground truth data and description file is freely available for downloading. MLDB2013 is provided for research or academic purposes only. If you use this dataset please cite this page.
To download the reference images (.tgz), please click here (size = 4707446).
To download the test images (.tgz), please click here (size = 8065661).
To download the description file (.pdf), please click here.

Uncompressed archives size: 5.7 Mb for references and 11 Mb for tests.

contact details

* Michela Lecca   |  FBK-irst, Povo, via Sommarive 18, I-38123 Trento, Italy   |  e-mail: l e cc a(at) fbk . eu