Ronald Cook
Samuel Cook
CSS 451
Final Project Report
Our Problem:
Our problem was that we wanted to generate better images without taking as many primary samples. With the current Ray Tracer that we had been using, it did one random sample per pixel and that was it. This resulted in quite a bit of aliasing. In order to improve these images we had to take multiple samples of every pixel in the entire image, usually around eight or so, in order for the image to look good. This makes generating the image take much longer to generate. We decided to implement Adaptive Super Sampling and Multi-Pixel Filtering in the hopes that we could generate better looking images for less computational power.
Our References:
Amortized Supersampling
http://www.cs.virginia.edu/~gfx/pubs/Yang_2009_AMS/yang2009.pdf
Ray Tracing With Adaptive Supersampling in Object Space.
http://www.cs.uaf.edu/~genetti/Research/Papers/GI93/GI.html
Selective and Adaptive Supersampling for Real-Time Ray Tracing
ftp://grmanet.sogang.ac.kr/pub/ihm/webpapers/09HPG.pdf
Supersampling.
http://en.wikipedia.org/wiki/Supersampling
Results Multi-Pixel Filtering (MPF):
Triangle Test:
Without MPF it took 8.516 seconds to generate an 800x600 image with 1 primary sample. With just the MPF enabled, it yielded an image that we would argue is actually better than the original and it only took 9.304 seconds. With 8 primary samples and no filtering it took 22.344 seconds, and 4 samples and filtering it took 14.476 seconds. The difference between the two images is that the one with 8 samples has sharper edges, but aliasing is still visible. With 4 samples and filtering the aliasing is much less noticeable, but it has softer edges. We think that this shows that if you use MPF to try to improve your image it will defiantly make the image much better, but that depending on your preferences doesn't necessarily make up for taking more primary samples.
Reflection, Transparency, Shadow Test:
Without MPF it took 5.61 seconds to generate the image with 1 primary sample. With MPF it took 6.029. The image with MPF did look slightly better. When we upped the samples to 4 and 8 samples the time difference between the image without MPF and with MPF was still less than a second. One thing we did notice though, was that with 8 samples and no MPF the image looked better than with 8 samples and MPF. We think this is because when you attempt to smooth out, already smooth edges it makes the edges less smooth.
Adaptive Super Sampling(SS):
Triangle Test:
Without SS it took 8.516 seconds to generate an image with 1 sample, and 22.344 seconds to generate an image with 8 samples. With SS it took 18.473 seconds to generate an image with 1 pass (5 samples on every pixel), and 19.321 seconds to generate an image with up to 3 passes (max 110 samples on every pixel) with an average of 5.157878 samples per pixel. The first image without SS obviously looked way worse than the image with SS but it was also much faster. With 8 samples however, you can't tell the difference between it and the one with SS up to 3 passes. This is even more of an improvement when you look at 16 samples vs. up to 5 passes because taking 16 samples takes 39.402 seconds, and up to 5 passes takes 20.571 with an average of 5.330052 samples per pixel. With this test there is almost no difference between the two images until you zoom in, and even then it's only a very slight difference favoring the image with 16 samples. The 20 second improvement in render time makes up for this though.
Reflection, Transparency, Shadow Test:
These results almost identically mirrored the results for the Triangle Test. The only difference was that time improvement when dealing with reflection and transparency was even more noticeable between without SS and with SS. Without SS it took 70.484 seconds to render the image, and with SS it took 25.411 seconds and 5.03185 average samples per pixel. The image differences were almost negligible until zoomed in, when they slightly favored the image with 16 samples. The 55 second difference in render time easily makes up for that though