Fua, Pascal 
  
  
Figure 4: Three different frames from the karate motion (set 30 frames apart), seen from various viewpoints. 
  
Figure 5: Virtual actor performing one of the recovered motions. 
determined by stereo correspondence, which we then use for reconstructing the 3-D marker. To this 3-D marker, we 
then also associate the 2-D coordinates from the remaining camera views onto which the 3-D coordinates of the marker 
projected correctly. Given that the visibility test has detected an erroneous 3-D reconstruction, we choose one of the 2-D 
coordinates computed via 3-D to 2-D projection, and calculate new 3-D coordinates. We then perform a new visibility 
test, and if this fails, we repeat the entire procedure. 
Occlusion check Once a 3-D marker has passed the visibility test, it needs to undergo the occlusion check: We want to 
ensure that the 3-D marker is not occluded from some camera views by another body part. To this end, we approximate 
body parts by solids, cylinders for limbs and a sphere for the head. In the case of limbs, the cylinder's axis is the 
corresponding bone and the radius is the average joint-to-marker distance of the markers associated to this body part. In 
the case of the sphere, the center is the mid-point of the segment. For each 3-D marker, a line is traced from the marker 
to the position of the camera, and tested for intersection with all body part solids. In case of intersection with a solid, 
the marker is most likely occluded from this camera view. Therefore, we conclude that we have used erroneous 2-D 
coordinates for reconstruction. As before, we choose other 2-D coordinates and repeat the process. 
32.2 Marker inventory When all the markers have been reconstructed and tested, we can proceed with tracking and 
fitting. More specifically, for each frame, we perform 3-D reconstruction, tracking from the previous frame into the 
present one, identification of all markers, and finally, fitting of the skeleton-and-marker model to the observations. In 
order for the fitting to work correctly, all markers must be present in every frame. To ensure this, we carry out a marker 
inventory after 3-D reconstruction and before fitting. Say we have just performed 3-D reconstruction using the 2-D data 
of frame [f], and we have thus obtained a set of markers. We then proceed with the following checks: 
1. If the number of markers reconstructed using trinocular stereo is smaller than the actual number of markers worn by 
the actor, we perform binocular reconstruction and add the newly calculated coordinates to the already existing list 
of markers. 
  
258 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 
  
3.3