RGB-NIR Scene Dataset
This dataset consists of 477 images in 9 categories captured in RGB and Near-infrared (NIR). The images were captured using separate exposures from modified SLR cameras, using visible and NIR filters. For more info on NIR photography, see the references below. The scene categories are: country, field, forest, indoor, mountain, oldbuilding, street, urban, water.
Before downloading the full dataset, you might want to look at the smaller dataset browser (36Mb) which is useful for exploring the data.
Please contact us if you require higher resolution or RAW images.
- RGB-NIR Scene Data (1Gb). TIFF images at 1024x768 resolution. Processed and aligned as described in the Technical Details section below.
The images were captured using Nikon D90 and Canon T1i cameras, using B+W 486 (visible) and 093 (NIR) filters. The cutoff between the two filters is approximately 750nm. After capture, the images were processed using dcraw. The colour capture was white balanced (dcraw -a). The NIR capture was processed using equal weights per band (dcraw -r 1 1 1 1), followed by averaging of the channels. The images were registered by extracting SIFT features at approximately 1500x2000 resolution (50% scale) and using RANSAC to find a similarity transform. The final transformation was recomputed via least squares from the inliers and used to resample both images in a common coordinate frame.
Thanks to the following people for their contributions to this database: Pierre-francois Laquerre, Nicolas Etienne, Noemie Vetterli, Caroline Duplain, Albrecht Linder, Sabine Süsstrunk.
- Multispectral SIFT for Scene Category Recognition. M. Brown and S. Süsstrunk.
International Conference on Computer Vision and Pattern Recognition (CVPR11) (pdf | bib)