Download Link
Baidu Yun Pan, http://pan.baidu.com/s/1ntl64mX
Google Driver, https://drive.google.com/file/d/0B176otcbAxRIQ05IekhZa3ltUFE/view?usp=sharing

PKU Person Orientation dataset: This dataset contains 4 video sequences with 976, 732, 1031, 855 frames respectively. The video sequences are captured in the campus surveillance scenes with 25 fps containing eight discrete orientations. The person images, orientation labels and position information are all manually annotated.

Each folder contains the video sequence data of one person. Label data is provided in the label.mat file. Position information is provided in the pos.txt file.

label.mat file structure: each row contains a label for the corresponding sample, the label-orientation matchup is shown as follow, 1(E) 2(NE) 3(N) 4(NW) 5(W) 6(SW) 7(S) 8(SE).

pos.txt file structure: each row contains the position infomation for the corresponding sample, the four values from left to right mean [x,y,width,height] and [x,y] is the left-top point coordinate of person image.

If you use this dataset please cite: @inproceedings{PKU_Orientation_2015_01, author = {Hong Liu, Liqian Ma}, title = {Oline person orientation estimation based on classifier update}, booktitle = {Image Processing (ICIP), 2015 22th IEEE International Conference on}, volume = {}, number = {}, year = 2015, month = Sep, pages = {}, address = {Quebec, Canada}, organization={IEEE} }