"Putting Objects in Perspective" Database
 
D. Hoiem, A.A. Efros, and M. Hebert.  "Putting Objects in Perspective", CVPR 2006.

This README describes both the validation set and the test set.

The images for this dataset are a resized (to maximum of 600x800) and 
renamed (so that they can be put in one directory) from the LabelMe database.  
Note that providing accurate and consistent ground truth in realistic scenes 
where objects are often hard to find and occluded is difficult.  I did my 
best, but there may be some mistakes which will lead to a slightly 
pessimistic quantitative analysis of a detector's performance. 

TESTSET:

The test set contains 600 images, but a subset of 422 of these images were 
used for experimentation (166 were removed because they were indoor, and 12 
removed because the labels were bad and couldn't be changed through the 
LabelMe tool).  

popTestImages - all test images
popTestset - contains the ts structure with ground truth and info on images
	ts.imnames{1:600} - the name of each image
	ts.gtruth(1:600) - the ground truth of each image
		bbox(nobj, 1:4) in format [x1 x2 y1 y2]
		objType(nobj) - 1 for car, 2 for ped
	ts.good_inds(1:422) - the indices of the images that are used
	ts.db(1:600) - information about the folder/names/annotations of the
			   images in the original LabelMe database 
popTestHorizons - contains horizon ground truth for the first 100 images
	horizons(1:100).imname - name of image
	horizons(1:100).b - y-position (from top in pixels) of horizon line
	horizons(1:100).height - height of image in pixels


VALSET:

Same as testset but with "Test" replaced by "Val" and "ts' replaced by "vs".