Illuminus D tasks

[ludimation]

[ technical ]

• compile build (see instructions in Nick's notes document)
• wish list
  • show training data as an overlay on the image
  • scale image interface so images can be re-proportioned to fit the data better
  • blow away the training on an image and re-train if necessary to put in fresh data for that image
  • input custom skeletal information (for flow calculations)
  • calculate target skeletal information based on joint momentum (both from the current skeleton, and the average of the kinect data)

[ visual ]

• color stylization (maybe just use drawings and no photographs—we'd need tons more if we'd like to go this route) — believe Rose might be working on some of these

[ divine needs ]

• provide figure drawings for use with the piece

[ deliverable — video ]

• screen-capture some video of the piece running for presentation

[ludimation]

[ done ]

• multiple user input (multiple harlequins on-screen at the same time)
• make data image name-labeled so we can just read all of the images in a directory and include them in the data set
• display images relative to user position & scale on the screen (can test with visual skeleton + depth image over-draw)
• save training data into the data directory so system does not have to be trained every time it is built (still have to test if it's loading properly)
• use lighter background for interactive mode (darker BG for debug mode still needs to be debugged)

[ludimation]

[ done ]

• debug image placement in virtual space so it matches bone data
• draw bones proportionally to depth data

[ludimation]

[ in progress ]

• load more images (memory issues)
• re-train system
• wish list
  • do not display the same image more than once every 3 frames (figure out a way to filter results returned — is it possible during the "predict" command? Or maybe during training model to data setting?)
  • slow down frame rate to 30, or even as low as 12 or so?
  • debug relative joint positions (store in separate data set so we can compare for which works best)
  • place relative to either joinPos - centerPos, or jointPos - torso components
  • see if there's a way to figure out where the ground is so that overall height of the pose relative to the height of the user can be used in the training data set
  • normalize data? (might be something that's already being done by GRT)
  • calculate axis angles for each joint based on head position and store in another data set?
  • user on-screen but no longer recognized bug
  • encrypt filenames into image labels (decimal representations of binary ASCII codes?) so that the number of images in the directory can change (will have to filter out images that are no longer around from joint data so system doesn't try to display them)

[ludimation]

Now that Illuminus D is over and others are going to be contributing code, I'm breaking this issue up into separate issues so that they can be addressed individually.