The database consists of FIR images collected from a vehicle driven in outdoors urban scenarios. Images were acquired with an Indigo Omega imager, with a resolution of 164x129 pixels, a grey-level scale of 14 bits, and focal length of 318 pixels. The camera was mounted on the exterior of the vehicle, to avoid infrared filtering of the windshield. Recorded images were manually annotated, where each pedestrian is labelled as a bounding box. To prevent bias introduced by border artifacts their height is subsequently upscaled by 5%. The pedestrians appear in an up-right position.
The dataset is divided in two: (i) Classification dataset: positives and randomly sampled negatives with a fixed height-width ratio of (1/2) and rescaled to 64x32 pixels, and (ii) Detection Dataset: Original positive and negative images with annotations.
Note: Only upright persons, with height over 10 pixels are annotated. Annotations may not be 100%right; in some exceptional cases, parts of the pedestrians may fall outside of the bounding box. Partially occluded pedestrians, or pedestrians not entirely contained inside the image are not labeled. The images were acquired in sequences thus, eventually, two consecutive images may in fact be the same.
The detection dataset was acquired in 13 different sessions, each containing a varying number of images. It comprises 15224 14 bit one channel images, with dimension 164x129 pixels. The Train set contains 6159 images, and the Test set contains 9065 images. Folders 'Train' and 'Test' correspond, respectively, to original training and testing images. Both folders have one subfolder for each independent sequence and a folder for the annotations files. These annotation files are compatible with the Pascal Challenge format. It may be useful to download their MATLAB development kit.
The classification Database is divided in a Train and a Test subset. The Train set contains 10208 positives and 43390 negatives, while the Test set contains 5944 positives and 22050 negatives. The annotated bounding boxes are resized to a constant aspect ratio (w/h) = 0.5 by changing their width appropriately. Any bounding box below 10 pixels in height is ignored. The remaining bounding boxes are resized to 64x32 pixels using bilinear interpolation. The negative samples were randomly selected from images not containing pedestrians.
If you make use of the LSI FIR database, please cite the following reference in any publication:
D. Olmeda, C. Premebida, U. Nunes, J.M. Armingol and A. de la Escalera. Pedestrian Classification and Detection in Far Infrared Images. Integrated Computer-Aided Engineering 20 (2013) 347–360
THIS DATA SET IS PROVIDED "AS IS" AND WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
The database on this page is copyright by us and published under the Attribution-NonCommercial-NoDerivs 3.0 Unported License. This means that you must attribute the work in the manner specified by the authors, you may not use this work for commercial purposes and you may not create derivative works. If you are interested in altering this work building upon it, or in a commercial usage you can contact us for further options.
For any questions, comments or other issues please contact Daniel Olmeda.