CIS565-Fall-2022 · horo-ursa · Oct 17, 2022 · Oct 21, 2022 · Oct 22, 2022 · Oct 22, 2022
diff --git a/README.md b/README.md
@@ -3,11 +3,77 @@ CUDA Denoiser For CUDA Path Tracer
 
 **University of Pennsylvania, CIS 565: GPU Programming and Architecture, Project 4**
 
-* (TODO) YOUR NAME HERE
-* Tested on: (TODO) Windows 22, i7-2222 @ 2.22GHz 22GB, GTX 222 222MB (Moore 2222 Lab)
+* Shixuan Fang
+* Tested on: Windows 11, i7-12700kf, RTX 3080Ti
 
-### (TODO: Your README)
+# Project Overview
 
-*DO NOT* leave the README to the last minute! It is a crucial part of the
-project, and we will not be able to grade you without a good README.
+| Original      | Denoised      |
+| ------------- |:-------------:|
+|<img src="img/noise.png" width=300 height=300>|<img src="img/10spp.png" width=300 height=300>|
 
+In this project, I implemented a pathtracing denoiser using A-trous wavelet filter.
+
+<img src="https://user-images.githubusercontent.com/54868517/197315297-c679c20d-7fa1-4edb-a3fa-af63aa814d58.png" width=700 height=250>
+
+A-trous is an approximation of gaussian blur by iteratively applying sparse blurs of increasing size, which significantly reduce the number of pixel read.
+
+## G-Buffer
+
+In order to use A-trous filter to denoise the result while preseving edges, we need extra information other than the path traced image. 
+
+|Normal|Position|
+|------|:-------:|
+|<img src="img/normal.jpg" width=300 height=300> |<img src="img/pos.jpg" width=300 height=300> |
+
+## Performance Analysis
+
+- how much time denoising adds to renders
+
+![800](https://user-images.githubusercontent.com/54868517/197315609-766b58ce-e6d0-4e43-9315-0e191bcae323.jpg)
+
+As seen in this figure, denoising only adds 2-3ms to the render, which is considered good enough for real-time rendering.
+
+- how denoising influences the number of iterations needed to get an "acceptably smooth" result
+
+|2spp|10spp|Ground Truth|
+|------|:-------:|:-------------:|
+|<img src="img/2spp.png" width=300 height=300> |<img src="img/10spp.png" width=300 height=300> |<img src="img/truch.png" width=300 height=300> |
+
+As seen in this table, 10spp with denoising is acceptably smooth compared to the ground truth(5000spp).
+
+- how denoising at different resolutions impacts runtime
+
+![image](https://user-images.githubusercontent.com/54868517/197312998-79594b67-bc01-4701-b9e7-e6a04f5094f3.png)
+
+As seen in this chart, with the increasement of screen resolution, the runtime of denoising also increases. This makes sense as there are more pixels to be considered, and A-trous filter needs more iterations to finish the whole picture, therefore denoising takes more time.
+
+- how varying filter sizes affect performance
+
+![image](https://user-images.githubusercontent.com/54868517/197314664-59f5fd76-cf51-473b-9c6d-76e3ab028a6c.png)
+
+As seen in this chart, when filter size increases the runtime of denoising also increases. This also caused by the same reason, which is with filter size increases, the A-trous filter needs more iterations, therefore increases denoising time.
+
+- how visual results vary with filter size
+
+|1|16|32|100|
+|------|:-------:|:--------:|:--------:|
+|<img src="img/1.jpg" width=200 height=200> |<img src="img/16.jpg" width=200 height=200> |<img src="img/32.jpg" width=200 height=200> |<img src="img/80.jpg" width=200 height=200> |
+
+This table shows how the denoised image looks with 1/16/32/80 filter size. The visual quality does increase from 1 to 16 and 16 to 32, but from 32 to 100 there aren't too much difference.
+
+- how effective/ineffective is this method with different material types
+
+|diffuse|specular|
+|------|:-------:|
+|<img src="img/diffuse.png" width=300 height=300> |<img src="img/10spp.png" width=300 height=300>|
+
+As seen in this table, A-trous works better with diffuse object in my opinion, as with specular objects it will blur the reflection face, which will result in a very blury surface. This is because position buffer and normal buffer can't really help split the edges on the same surface.
+
+- how do results compare across different scenes 
+
+|cornell|cornell_ceiling|
+|------|:-------:|
+|<img src="img/cornell.png" width=300 height=300> |<img src="img/10spp.png" width=300 height=300>|
+
+As seen in this table, with same parameters, cornell_ceiling scene looks better and smoother than the cornell scene. This is because it has a larger light source, therefore making path tracing converge faster than small light source. Better path traced image will definitely makes denoised image better.
diff --git a/img/1.jpg b/img/1.jpg
diff --git a/img/1000.jpg b/img/1000.jpg
diff --git a/img/10spp.png b/img/10spp.png
diff --git a/img/1200.jpg b/img/1200.jpg
diff --git a/img/16.jpg b/img/16.jpg
diff --git a/img/2spp.png b/img/2spp.png
diff --git a/img/32.jpg b/img/32.jpg
diff --git a/img/80.jpg b/img/80.jpg
diff --git a/img/800.jpg b/img/800.jpg
diff --git a/img/cornell.png b/img/cornell.png
diff --git a/img/diffuse.png b/img/diffuse.png
diff --git a/img/gauss.png b/img/gauss.png
diff --git a/img/noise.png b/img/noise.png
diff --git a/img/normal.jpg b/img/normal.jpg
diff --git a/img/pos.jpg b/img/pos.jpg
diff --git a/img/truch.png b/img/truch.png
diff --git a/scenes/cornell_ceiling_light.txt b/scenes/cornell_ceiling_light.txt
@@ -48,6 +48,16 @@ REFR        0
 REFRIOR     0
 EMITTANCE   0
 
+// Diffuse blue
+MATERIAL 5
+RGB         .35 .35 .85
+SPECEX      0
+SPECRGB     0 0 0
+REFL        0
+REFR        0
+REFRIOR     0
+EMITTANCE   0
+
 // Camera
 CAMERA
 RES         800 800

diff --git a/src/main.cpp b/src/main.cpp
@@ -30,6 +30,7 @@ float ui_normalWeight = 0.35f;
 float ui_positionWeight = 0.2f;
 bool ui_saveAndExit = false;
 
+
 static bool camchanged = true;
 static float dtheta = 0, dphi = 0;
 static glm::vec3 cammove;
@@ -70,6 +71,7 @@ int main(int argc, char** argv) {
     height = cam.resolution.y;
 
     ui_iterations = renderState->iterations;
+    ui_iterations = 5000;
     startupIterations = ui_iterations;
 
     glm::vec3 view = cam.view;
@@ -165,7 +167,11 @@ void runCuda() {
         pathtrace(frame, iteration);
     }
 
-    if (ui_showGbuffer) {
+    if (ui_denoise) {
+        Denoise_Image(ui_colorWeight, ui_normalWeight, ui_positionWeight, ui_filterSize);
+        showDenoisedImage(pbo_dptr, iteration);
+    }
+    else if (ui_showGbuffer) {
       showGBuffer(pbo_dptr);
     } else {
       showImage(pbo_dptr, iteration);

diff --git a/src/main.h b/src/main.h
@@ -42,6 +42,7 @@ extern float ui_normalWeight;
 extern float ui_positionWeight;
 extern bool ui_saveAndExit;
 
+
 void runCuda();
 void keyCallback(GLFWwindow *window, int key, int scancode, int action, int mods);
 void mousePositionCallback(GLFWwindow* window, double xpos, double ypos);