Dense 3D Reconstruction and Tracking of Dynamic Surface - Advances in Image and Graphics Technologies - page 219

obtain dense and smoothing surface, the accurate location of the point corresponding
with the left point can be achieved by bilinear interpolation according to the distance
from that point to the four neighboring block centers in the image I pl . Similarly, when
we perform the tracking between the temporal images, we can first get four values by
shifting the point with the four shift parameters of the neighboring blocks, and then
the four values can be interpolated to obtain the accurate tracking location.

3 Results and Evaluation

In the previous sections, we show how our method can be applied to the
reconstructing and tracking of dynamic surface. In the experiment, first, simulated
data is used to evaluate the performance of the algorithm. Then,

(a) Error of image points (b) Error of reconstructed 3D points
Fig. 3. Error on the ground truth,(a) is the RMSE curve of image points on the ground truth,(b)
is the RMSE curve of the reconstructed 3D points on the ground truth. The top curve of the two
figures is the RMSE for the matching without block overlapping, and the bottom curve is that
of with 50% block overlapping matching.

We validate our approach using real video data acquired by two calibrated cameras
with 25fps to demonstrate that it can produce good results for very different kinds of
materials, and at the same time, motion field is easily generated by the tracking
information.

3.1 Simulated Experiment

To evaluate the performance of our algorithm, the simulated data generated by 3D
max is used as ground truth for our first experiment. We produced a stereo image
sequence of 100 frames with 800×600 image size in which a piece of cloth is
fluttering with different shapes in the breeze, and there is a small movement in
previous 20 frames and a sharp movement from 35th frame to 45th frame in the stereo
image sequence. At the same time, 400 pairs of vertices of the left and right image
sequence and the corresponding spatial 3D points are generated as the benchmark to
evaluate the performance of our algorithm.