In many cases, distinct words contain errors. However, there are cases where distinct words contain a unique motion, which is not similar to motions found in the data, but does not contain errors.
We deal with this challenge by defining smaller motion-words using sub-groups of joints, such as the left and the right side of the body, and the top and bottom of the body. Moreover, instead of searching in a single motion sequence, we define a larger-context to search for the KNN that consists of related motion capture sequences.
Searching in larger-scale context instead of just in the same sequence, and using joint groups analysis, we reduce the possibility of mistaking unique motions as ones that contain errors.