Summer 2003 Independent Study (Cleaning of Motion Capture Data)
This summer's 950 was dedicated to cleaning motion capture data of 5 dogs that was captured during Autumn Quarter of 2003. The dogs were named Stevie, Tutti, Bailey, Coaly, and Madelaine.
There are different ways to clean motion capture data depending on what equipment you are working with and the software you have purchased. In Spring of 2002, I learned to first clean the motion capture data in Vicon's Workstation, then take it into House of Move's Diva for further cleaning. Then, the data would be further manipulated in Aliaswavefront's Maya.
ACCAD has recently replaced Diva with Vicon Motion Builder for their pipeline. I've seen this program demonstrated, but don't know it yet.
This summer, I was able to clean all the files in Workstation and plan on taking them into Diva over break and during Autumn Quarter. Each file took approximately 10 hours with the stair climbing data taking a little longer.
In Workstation, the video data (.tvd files) are reconstructed into .c3d format. The c3d format is essentially unlabelled dots in space. Example of unlabeled marker data. I had to hand label all of these.
When reconstructing tvd data within Workstation, you can tell it how many cameras must see a marker for it's trajectory to be recognized and show up. I had to set this to only 1 camera, but usually it is set as 3. I had to set it to 1 camera because I had so many gaps in my data where markers disappeared from view and weren't registered. The lower the number of cameras you reconstruct with, the more "ghost" trajectories you might have. I would rather clean a bunch of garbage out of my files than miss a marker's data all together.
A ghost trajectory is usually a very short lived duplicate of another marker. View an example of a ghost marker highlighted in yellow.
For a file that should have only had 66 markers/trajectories, in one instance, there were 304. I had to label all of these and discern what was garbage and what needed to be kept. See an example of how short and broken my trajectories were.
Once labelled, I had to fix swaps and spikes within the trajectories. A swap is when one trajectory gets confused with another within the computer. For example, view a movie of a swap in which a yellow marker starts out as the back of the elbow on the dog, then midway, becomes the front of the elbow.
A spike is when the marker pops out of place and then back into place. View a movie of a spike in which a blue marker pops out of place, then back. If these spikes are not taken out before creating a skeleton within Diva, the limb may travel believably then all of a sudden move off at an unnatural angle.
Some markers had both swaps and spikes. View a movie of a swap then a spike of a yellow elbow marker before it disappears. Data like this was very difficult and tedious to clean. I had to highlight one or two markers (e.g. of the elbow or knee) and quite literally go frame by frame watching them constantly. The elbows, knees, and feet gave me the most problems on the dogs.
Also, right before a marker's trajectory disappears, the "tail" data is almost always bad. I had to cut these tails out of the trajectories. View a movie of bad tail data in which a yellow marker flies out of position and off into space before disappearing.
Short gaps of 10 to 20 frames can be filled within Workstation usually without bad effects. Longer gaps have to be filled within Diva or Motion Builder. I tried copying a pattern from a similar marker to fill the gaps in one marker, but this didn't give believable or clean results.
View a list of files I cleaned. They are listed by the dog's name and type. These link to movies of the cleaned Workstation data.