Group 10: Mid-Term Report
Our test-reel so far is here:
Preview3.avi
Or here with different compression...
Web.avi
So far we've been tightening our grip on what has been done, what can be done, and how we can apply it to our project.
The scenes were first attempted using standard point morphing on single layered frames. This had a look known in the professional world as "really bad". Extracting the foreground object and motion gives the scene a much more realistic look due to eliminating the object/object occlusion problem which creates most of the large visual artefacts in the morph. Also the features don't need to be accurate - heightening the chance for automatic simple feature detection. The scenes shown immediately below were the two images used to generate the camera rotation morph. This is not a view-morph...this is only one of the steps. There was no camera intrinsic calibration or measurements done. The background plate morph was tricky due to the corners of the image not blending in fast enough on one side and too fast on the other. It looked better for the blend to happen almost immediately on the left (the corner being revealed) and to blend very late on the right corner. There were two different interpolations done and then they were blended in the middle behind front layer of the jumping goofy guy. Most of this tweaking will be drastically reduced by employing view-morphing instead of regular feature-based image morphing.
Here's a couple scoped and greened frames from the rotation morph. From camera A:
And B:
Slow Motion
For slow-motion the frames were interpolated on a single camera, but there's no reason we can't interpolate between morph-synthesized frames to get a slow-motion effect while virtually rotating the camera.
Here's the initial two shots.
Camera A-Frame7:
Camera A-Frame8:
Rotoscoping is key here, especially since right now these are just feature-based morphs. We will be adding the view-morphing addition to feature-based morphing in our next phase. Here's the scoped and greened frames plus three interpolated frames in between. The effect was quite effective up to about 5 interpolated frames for every 1 real frame (1/5 real frames for a significant amount of time) and at this point there was a slight repetitive shudder to the image. In the test-composite you can see this problem near the top of his jump. His jump is a variable increasing interpolation - time slows down incrementally over the jump, peaking at 1/7 real frame ratio at the top of his jump. There are also some morphing artefacts around the perimeter of the image. These are a result of morphing that is eliminated with the view-morphing addition and most of it will disappear.
This is the scoped and greened shot with the feature base.
This is a real shot from Camera A:
Synthesized shot 1:
Synthesized shot 2:
Synthesized shot 3:
This is a real shot from Camera A:
Here's the test-composite for this shoot.
Preview3.avi
Or here with different compression...
Web.avi
Links to this Page
- DVFX2003: Groups last edited on 1 May 2003 at 4:53 pm by davinci.cc.gatech.edu
- 4grounD last edited on 19 March 2003 at 6:17 pm by adsl-33-184-122.asm.bellsouth.net