Segmentation and tracking in a general setting is hard. We learned this the hard way. Initially, we had considered the silhouette extraction procedure to be straight-forward and relatively simple. As we analyzed our diving photage, this quickly became untrue and quite challenging. Much of our efforts, while initially intended towards disparity calculation between divers, were redirected towards segmenting and tracking the diver. While more advanced techniques exist for segmentation, we feel that our choices for segmentation are ideal for the time frame at hand.

Designing a simpler (yet effective) segmentation and tracking procedure gave us the freedom to explore certain disparity techniques that wouldn't have been possible had we spent the time to evaluate a richer segmentation scheme. We implemented blob/centroid comparison as well as a "cardboard" parameterized model to capture the leg and arm motions of the dive.

Some extensions to this current analysis would include a better model of the diver, perhaps a 3-dimensional model to capture the twists and turns that are thrown away in our current system. Running the algorithms in real time would also be a good extension. This would be a simple extension, but would be a step towards providing an objective analysis tool for diving judges. Also, an initial silhouette must be manually set for a given movie clip. We would like to automate this process with perhaps some a priori knowledge about dive positions, skin values, lighting, etc. Finally, we would like to augment our system to handle two dives in the same video photage, with inter-occlusions -- as would be the case in olympic diving photage.