The tracking algorithm was tested in the following manipulation scenarios:
The experiments shown below were performed with ground truth markers on the objects, however, the tracking results shown do not use the marker information. The tracker used point clouds from a single Asus Xtion Pro Live camera (equivalent to the MS Kinect.) Scenarios 1 and 2 (human manipulating rope and cloth) used color information. Scenario 3 (robot manipulating rope) uses no color information. All videos for each scenario were generated with the same set of parameters for the tracker.
The left column of videos below shows the camera image of the object being tracked. The right column is a rendering of our state estimate when using our tracking algorithm.
Full videos for all of the test sequences referred to in the paper are included below (1X speed)
Human Manipulating Cloth |
|||
Experiment Task |
|
||
Fold diagonal left |
Download |
||
Fold diagonal right |
Download |
||
Fold double |
Download |
||
Fold long |
Download |
||
Fold short |
Download |
||
Fold triple |
Download |
||
|
For the following videos of a human manipulating rope, the point cloud is also drawn in the state estimate rendering. The points are colored in a range between green and black, where darker points are more likely to be outliers. |
Human Manipulating Rope |
|||
Experiment Task |
|
||
Overhand knot (I) |
Download |
||
Overhand knot (II)The tracking for this task fails because an occlusion causes an ambiguity when one end of the rope is moved under another section of the rope (0:15). However, this ambiguity doesn't arise anymore when two cameras are used (look at the comparisons section). |
Download |
||
Double overhand knot |
Download |
||
Figure-eight knot (I) |
Download |
||
Figure-eight knot (II) |
Download |
||
Overhand knot + untieIn this task, the rope gets tied in a way such that untying the rope requires just to pull both ends of it. |
Download |
||
Robot Manipulating Rope |
|||
Experiment Task |
|
||
Overhand knot |
Download |
||
Figure-eight knot (I) |
Download |
||
Figure-eight knot (II) |
Download |
||
ComparisonsIn this section we compare the performance of our tracking algorithm as we use more cameras or no color information. |
Experiment Task |
|
||
Overhand knot (I):
|
Download |
||
Overhand knot (II):
|
Download |
Other deformable objectsIn this section we present qualitative results of our algorithm tracking other objects. |
Experiment Task |
|
||
Sponge manipulation:
|
Download |