We provide hand data sets used in the following paper:
Quan Yuan, Ashwin Thangali, Vitaly Ablavsky and Stan Sclaroff ,
"Multiplicative
Kernels: Object Detection, Segmentation and Pose Estimation,"
CVPR, 2008.
part 1, part 2, part 3. To get annotations, in matlab call "info = imfinfo(tif_filename);"
In the string "info", the first four numbers define a bounding box, and the last two numbers are raw angles of \theta_1 and \theta_2. \theta_1 is in degrees while \theta_2 is in radians.
The two raw angle parameters can be normalized to [0,90] by following two commands in Matlab:
all_parameters(:,1) = (all_parameters(:,1) - 90); all_parameters(:,2) = (all_parameters(:,2)*180/pi + 90);
The dataset has two parts, A) 902 hand images from expert signers, with silhouettes. B) 154 hand images from non-expert signers, with bounding boxes. Each hand image is also rotated by [-40, 20 ,0 ,20 ,40] degrees to generate the training and test sets in the CVPR08 paper.