
Image databases for Hierarchical Models of Scenes, Objects, & Parts
*******************************************************************

These images, and corresponding sparse features, were used to run
the experiments reported in the following journal publication:
  Describing Visual Scenes Using Transformed Objects and Parts
  E. B. Sudderth, A. Torralba, W. T. Freeman, and A. S. Willsky
  Int. J. Comp. Vision, vol. 77, pp. 291-330, 2008.

Images are also available through LabelMe:  http://labelme.csail.mit.edu/

Directory imagesSharingIJCV contains the dataset of 16 object categories,
while directories office and streetFixScale contain office and street scenes.

In addition to the raw images, we include Matlab .mat files containing 
precomputed features, stored in a struct.  For the code which generated these
features, see scripts/databaseSharing and scripts/databaseScenes.

For the 16-category dataset, the struct encodes feature jj in image ii as follows:
* dictionary -> vector quantized SIFT descriptors defining "visual words"
* appearance{ii}(jj) -> maps feature jj in image ii to a visual word
* position{ii}(:,jj) -> 5-dim. descriptor of feature ellipse (first two entries are x,y position)
* classes(ii) -> object category for image (zero for background)
* files{ii}   -> filename for image (matches LabelMe database) 
* images{ii}  -> rescaled copy of image 

For the scene dataset, the struct encodes feature jj in image ii as follows:
* dictionary -> vector quantized SIFT descriptors defining "visual words"
* appearance{ii}(jj) -> maps feature jj in image ii to a visual word
* descriptor_left{ii}(:,jj) -> 128-dim. SIFT descriptor 
* shapeROI_left{ii}(:,jj) -> 5-dim. descriptor of feature ellipse (first two entries are x,y position)
* typeROI_left{ii}(jj) -> Integer indexing type of ROI 
* classes{ii}(jj)   -> Object class label (integer) 
* instance{ii}(jj)  -> Object instance (polygon) label 
* overlap{ii}(:,jj) -> Measure of ROI overlap with ground truth segmentation
* files{ii} -> filename for image (matches LabelMe database) 

