Here you can find the benchmark dataset from the paper:
The dataset contains three sequences in which the same articulated object is manipulated. The 3D model information of this Kubito-object can be found here: Blender and Wavefront OBJ files.
The following calibration info corresponds to the rectified sequences (note these are slightly different from the Rigid Pose dataset):
focal_length = 583.9102; % (in pixels)
baseline = 70.8646; % (in mm)
nodal_point_x = 317.9841; % column (in pixels)
nodal_point_y = 277.0105; % row (in pixels)
Pixels are square and focal lengths and nodal point are identical in both (rectified) images.
See the following previews to get a better idea of the sequences: