081217 / Reprojection

previous | next

To avoid any confusion, the following screens don't represent at all what Atom is going to look like. Rather just a single rule L-system colored by a function of eye relative position that I use for testing. I have the color as a function of eye relative position to represent worst case, changing of color as I fly through the scene. If motion and fast changing color can be solved, then the solution is robust.

Inputs

Only a small number of points get re-computed per frame. Also given my visibility solution, many of these points are not actually visible. Points are sparse around edges of geometry also do to the stochastic visibility solution. However sparse distribution around the edges of the screen is do to the fisheye projection. Fisheye edges will get some kind of blur to hide artifacts in final engine.

Points updated per frame, view 1280x720 version.

Convergence

The result of recirculating the stochastic visibility solution over time is a nicely detailed and anti-aliased view of the scene. This solution depends on blending in new points which are not occluded. Converging to this level of quality can happen in 32 frames (one second).

Convergence 1, view 1280x720 version.

Convergence 2, view 1280x720 version.

Effects

Really hard to show this without a video which I don't have time to make. Playing with the reprojection position, feedback, and blur amount, can be used to do all sorts of interesting stuff. Such as having the fluid like interactions between the geometry.

Fluid effect, view 1024x512 version.

The blur effect below isn't actually a real blur, but rather an adjustment of the feedback position. The dots in the image are from my lazyness at catching NANs.

Blur effect, view 1024x512 version.

Issues

The right trade off has to be found between getting high quality sharpness + anti-aliasing and fast convergence during motion and color changes. Still working on finding the right balance. Going for fast convergence alone doesn't yield the best results...

Dithered look + fluid effect with very fast convergance, view 1280x720 version.

Great GI Papers

SSDO - Screen space global illumination. Surprised that it took this long for this kind of paper (screen space GI) to get released. Results are IMO good for fine details, but not too thrilled at larger low frequency results. Might be neat to combined this technique (for detail GI) with the technique in the second paper.

ISM - Instant radiosity via imperfect shadow maps (render point proxy geometry into shadow maps and hole fill). Very inspiring results.