Thanks! BTW, I was holding the camera! Everything but the pillow in the FG and the figure was a quicky handheld of my living room. It made the match-move rather tricky (and not 100% spot on), but that was the point of the project.
The pillow is part of the model I built in Blender, just a cube subdivided and sculpted. It roughly matches one that was in the live action that would have occluded anyone sitting on the couch, so I made it part of the CG to avoid any major rotoscope/matte work. It also made the match-move in Blender more complicated, because it’s so close to the camera that the parallax was pretty severe.
The major steps in the process were building a mockup of the room with basic primitives, using rough measurements of the room to get the scale close. I integrated the Voodoo camera move data into this setup, with some additional camera motion I basically just did by patient eyeballing. After getting a fair match move on the room model I added the pillow mesh and tweaked it into place, then added the Mancandy figure and started finessing the camera position & motions, and positions and scale of the model elements until everything started to fit tightly. I never got a 100% perfect match (I think a real lens and Blender’s camera have somewhat different perspective characteristics, even when the focal lengths are matched), but by disguising some of the less successful frames with the figure’s motion, the overall effect’s pretty solid.
I also had to add some meshes to act as “shadow casters and catchers”. These were on a separate Render Layer with special lighting to get a very high-contrast shadow pass, which I colorized and added to the other passes in the Compositor.
Lastly, to make sure the “render-perfect” finish of the CG didn’t look out of place, I rendered a pass to JPEG frames using a low Q factor (55), which gave it a noisy “grain,” then composited that in a separate .blend over the high-Q BMP frames, using an alpha sequence made from the original render layers to isolate only the CG elements. I tried adding a “noise-pass” in the original composite but couldn’t get a good match to the distinctive JPEG artifacts look.