CVPR 2008 – 3D Pose Refinement from Reflections – Paper ID 576
Supplemental material
We supply two videos in which we compare tracking results using either texture only or both texture and lighting information. In both cases, the tracking was done frame by frame. No temporal consistency constraints were imposed to better differentiate the intrinsic capabilities of the two approaches.
In this video, we track a very textured and shiny coffee box. Since it is cylindrical, different rotations around a horizontal axis yield very similar projections of its vertical part. As a result, this degree of freedom is poorly recovered using texture only, as evidenced by the fact that the projections of the top of the box are wrong. By constrast, when using the lighting information, these rotations are correctly recovered and the top of the box projects where it should.
This time we track a poorly textured shiny duck. In this case, there is not enough texture information and using it only results in jitter. Using the lighting information resolves this problem.