CS 184 Project 3-1 Writeup

Tianchen Liu 3034681144, Riley Peterlinz 3036754368

Link:
https://cal-cs184-student.github.io/project-webpages-sp23-CardiacMangoes/proj3-1/index.html

Overview

In Part 1, we first generate rays that shoots through the camera into the scene Then we calculate intersections between the rays and the primatives. We write methods to cacluate ray-triangle intersection and ray-sphere intersection.

Then in Part 2, we implement BVH optimization to speed up calculations of ray-primative intersections by preprocessing the primatives into bounding boxes.

Then in Part 3, we deal with illumination and implement direct lighting. We implement direct lighting in two ways: uniform hemisphere sampling and lighting sampling. As we can see, lighting sampling creates less noisy pictures.

For part 4, we implemented multiple bounces for indirect lighting. Then, for part 5 we used adaptive sampling to intelligently sample different parts of the image depending on the variance of their sampled distributions, leading to faster rendering times for high sampled scenes.

Part 1

Walk through the ray generation and primitive intersection parts of the rendering pipeline.

For each pixel on the screen, the camera generates a ray that “shoots through the pixel from the camera”.

For each location on the screen, we first imagine there’s a virtual camera sensor in the world, and each location on the screen would translate to a location/coordinate on the censor. Then we translate a pixel to that coordinate on the virtual censor. Then we generate a ray that starts from the camera position, shoots in the direction of coordinate in the virtual censor. The origin is the position of the camera, the direction vector is the vector that starts from origin and ends in the location of the corresponding location on the sensor (but normalized). We assign the max_t and the min_t.

The code to do it is as follows:

  // vec is direction vector
  auto vec = Vector3D(tan(PI/180*hFov/2)*(x-0.5)*2, 
                      tan(PI/180*vFov/2)*(y-0.5)*2, 
                      -1);
  vec.normalize();
  auto ray = Ray(pos, c2w*vec);
  ray.max_t = fClip;
  ray.min_t = nClip;
  return ray;

Now we know how to shoot a ray from the camera to a location on the vitual censor, for every pixel on the image space, we sample ns_aa rays that shoots through that pixel (or the equivalent to where that pixel would be in the virtual sensor) from the camera.

  int num_samples = ns_aa;          // total samples to evaluate
  Vector2D origin = Vector2D(x, y); // bottom left corner of the pixel
  auto sampler = UniformGridSampler2D();

  Vector3D ret;

  for (int i = 1; i <= num_samples; i++) {
    auto sample_point = origin + sampler.get_sample();
//    Vector2D normalized_sample_point = {sample_point.x / sampleBuffer.w, sample_point.y / sampleBuffer.h};
    auto ray = camera->generate_ray(sample_point.x / sampleBuffer.w, sample_point.y / sampleBuffer.h);
    ret += est_radiance_global_illumination(ray);
  }


  sampleBuffer.update_pixel(ret * (1.0/num_samples), x, y);
  sampleCountBuffer[x + y * sampleBuffer.w] = num_samples;

Explain the triangle intersection algorithm you implemented in your own words.

We first implement the Moller-Trumbore algorithm taught in lecture

tuple<double, double, double> MollerTrumbore(const Ray &r, Vector3D p1, Vector3D p2, Vector3D p3) {
  auto E1 = p1 - p3;
  auto E2 = p2 - p3;
  auto S = r.o - p3;
  auto S1 = cross(r.d, E2);
  auto S2 = cross(S, E1);
  auto coeff = 1.0 / dot(S1, E1);
  auto t = coeff * dot(S2, E2);
  auto b1 = coeff * dot(S1, S);
  auto b2 = coeff * dot(S2, r.d);
  return {t, b1, b2};
}

This algorithm takes in a ray and three vertices of the triangle and returns the t for the intersection of the ray and the triangle as well as the barycentric coordinates b1 and b2.

We first check the t returned is within min_t and max_t of the ray:

if (!(0 <= t && r.min_t <= t && t <= r.max_t)) {
    // Ray does not shine on plane
    return false;
  }

Now we check if the barycentric coordinates are valid (inside triangle):

if (!(0 <= b1 && b1 <= 1 && 0 <= b2 && b2 <= 1 && b1 + b2 <= 1)) {
    return false;
  }

If so, we set the max_t of all the ray to be t and assign values to the intersection accordingly before returning true.

  r.max_t = t;

  // The following is not needed if 
  isect->t = t;
  // surface normal
  isect->n = b1 * n1 + b2 * n2 + (1 - b1 - b2) * n3;
  isect->primitive = this;
  isect->bsdf = get_bsdf();

  
  return true;

Show images with normal shading for a few small .dae files.

Cow:

Beetle:

Teapot:

Part 2

Walk through your BVH construction algorithm. Explain the heuristic you chose for picking the splitting point.

We use the average centroid method to find the splitting point.

• We first compute the average centroid of all bounding boxes of primatives, - then we find the longest axis,
• Then we group primatives based on which side of the average centroid the primative is on along the longest axis

Here’s our code. See comments for explanation/walkthrough.

  BBox bbox;

  int cnt = 0;

  Vector3D average_centroid;

  // Compute average centroid:
  for (auto p = start; p != end; p++) {
    cnt++;
    BBox bb = (*p)->get_bbox();
    bbox.expand(bb);
    average_centroid += bb.centroid();
  }

  // If the number of primatives in this level of recursion is less than max_leaf_size, then this is the base case and we should return a BVHNode with null L and R pointers that corresponds to this bounding box and these primatives.
  if (end - start <= max_leaf_size) {
    auto *node = new BVHNode(bbox);
    node->start = start;
    node->end = end;
    return node;
  }

  // Decide what's the longest axis of the current bounding box.
  average_centroid /= cnt;
  double delta_x = bbox.extent.x;
  double delta_y = bbox.extent.y;
  double delta_z = bbox.extent.z;
  bool x_longest = (delta_x >= delta_y && delta_x >= delta_z);
  bool y_longest = (delta_y >= delta_x && delta_y >= delta_z);
  bool z_longest = (delta_z >= delta_x && delta_z >= delta_y);

  // If the primative is on the "<=" side, then put it in the left vector.
  // Otherwise put it in the right vector
  vector<Primitive *> left;
  vector<Primitive *> right;
  for (auto p = start; p != end; p++) {
    auto centroid = (*p)->get_bbox().centroid();
    if (x_longest && centroid.x <= average_centroid.x || y_longest && centroid.y <= average_centroid.y || z_longest && centroid.z <= average_centroid.z) {
      left.push_back((*p));
    } else {
      right.push_back((*p));
    }
  }
  
  // Reorder the primatives vector so that all left primatives come before right primatives.
  for (int i = 0; i < left.size(); i++) {
    *(start+i) = left[i];
  }
  for (int i = 0; i < right.size(); i++) {
    *(start + left.size() + i) = right[i];
  }

  // Assign start and end pointers of the node. The left node should be constructed with primatives in the left vector, and the right node should be constructed with primatives in the right vector.
  auto *node = new BVHNode(bbox);
  node->start = start;
  node->end = end;
  node->l = construct_bvh(start, start + left.size(), max_leaf_size);
  node->r = construct_bvh(start + left.size(), end, max_leaf_size);

  return node;

Show images with normal shading for a few large .dae files that you can only render with BVH acceleration.

Compare rendering times on a few scenes with moderately complex geometries with and without BVH acceleration. Present your results in a one-paragraph analysis.

The time it takes to collect primatives is roughly the same for both methods (around 0.0011 seconds). The time it takes to build BVH from primatives is, of course, longer for BVH, (0.0027 seconds for BVH vs 0.0001 seconds for naive solution). But this difference of time is negligible in the grand scheme of things. The time it takes to render is a lot longer for the naive method. It takes on average 63.8475 seconds without BVH acceleration and 0.0704 seconds with BVH acceleration. The average rays/second is of course, inverse to this ratio because the amount of rays stays the same. This is because the average intersection tests for a ray is drastically decreased. Without BVH acceleration it does 2.95974625 intersection tests per ray on average, but with BVH acceleration it does 1367.904394 intersection tests per ray on average.

Part 3

Walk through both implementations of the direct lighting function.
Show some images rendered with both implementations of the direct lighting function.

Hemisphere sampling:

We obtain the intersection between the ray and the primative first. Then we sample num_samples directions uniformly at random on the hemisphere in the direction of the normal of the intersection. We then generate a ray that starts from the intersection and shoots in this direction. We then obtain the emission from the new intersection (see if this reflection ray intersects a light source), and if so, the total should add the emission obtained from the camera from this ray and its reflection using this reflection equation: https://cs184.eecs.berkeley.edu/sp22/lecture/13-23/global-illumination-and-path-tra .The sample.z is the cos(theta) in this case. Then we divide the total summation of results by the number of samples.

  Vector3D result;
  for (int i = 1; i <= num_samples; i++) {
    auto sample = hemisphereSampler->get_sample();
    auto ray_after_bounce = Ray(hit_p, o2w * sample, r.max_t);
    ray_after_bounce.min_t = EPS_F;
    Intersection isect_after_bounce;
    bool intersected = bvh->intersect(ray_after_bounce, &isect_after_bounce);
    if (intersected) {
      result += sample.z * isect_after_bounce.bsdf->get_emission() * isect.bsdf->f(w_out, sample);
    }
  }
  return result / num_samples;

Importance sampling:

For importance sampling, we iterate over the lights in the scene and test intersection between our ray and the light. If there are no colisions, we sample from that light. This helps beecause now we are only sampling rays that would hit a light and thus the result has less noise since we are not sampling empty space that does not contribute light. For area lights we sample ns_area_light times, for point lights we only sample once since the ray to that light is the only ray that would hit from the point light. The amount of light being collected each sample is scaled by the bsdf and the angle between the light’s normal and the normal of the object at the point we are tracing. We also divide by the pdf of the sample for how important it is to the distribution.

for (auto light: scene->lights) {

      Vector3D wi; // direction between p and light source
      double distToLight; // distance between p and light source
      double pdf; // probability density function evaluated at the wi direction

      Ray ray_after_bounce;

      float costheta; // angle between light source's and the ray shooting from p to light source

      Vector3D L;

      for (int s = 0; s < ns_area_light; s++) {
          Vector3D sample = light->sample_L(hit_p, &wi, &distToLight, &pdf);

          wi.normalize();

          auto wi_object = w2o * wi;

          costheta = dot(wi, isect.n);

          ray_after_bounce = Ray(hit_p, wi, 1); // assigns the location hit_p and direction wi of the next ray

          ray_after_bounce.max_t = (double) distToLight - EPS_F;
          ray_after_bounce.min_t = EPS_F;

          Intersection isect_after_bounce;

          // if ray hits light, add luminance
          if (!(bvh->intersect(ray_after_bounce, &isect_after_bounce)) && costheta > 0) {
              L += sample * isect.bsdf->f(w_out, wi) * costheta / pdf;
          }

          if (light->is_delta_light()) {
              break;
          }
      }

      if (light->is_delta_light()) {
          L_out += L;
      } else {
          L_out += L / (ns_area_light);
      }

  }


  return L_out;

Importance:

Hemisphere:

Importance:

Hemisphere:

Importance:

Hemisphere:
N/A Because we can’t render point light source with Hemisphere sampling

Focus on one particular scene with at least one area light and compare the noise levels in soft shadows when rendering with 1, 4, 16, and 64 light rays (the -l flag) and with 1 sample per pixel (the -s flag) using light sampling, not uniform hemisphere sampling.

One light ray:

4 light ray:

16 light ray:

64 light ray:

The noise level decrease very noticably and the soft shadows gets much more realistic as the light rays amount goes up.

Compare the results between uniform hemisphere sampling and lighting sampling in a one-paragraph analysis.

First of all, uniform hemisphere sampling cannot render point light source. For area light source, hemisphere sampling generally produces a much more noisier image that’s somewhat darker because of the noise. The shadows for lighting sampling is smoother because of less noise as well. In general, lighting sampling creates a much more realistic image.

Part 4

For Indirect Lighting we let a ray of light continuously bounce around a scene. The zero bounce is any emissive material. The first bounce is “direct lighting” from the light source onto the objects in the scene. Any more bounces is “indirect lighting” that occurs from light that bounces off a material and still has energy to illuminate other materials. We can sum the direct emission with direct and indicet lighting like so:

Vector3D PathTracer::est_radiance_global_illumination(const Ray &r) {
  Intersection isect;
  Vector3D L_out;

    if (!bvh->intersect(r, &isect)) return L_out;

    L_out = zero_bounce_radiance(r, isect);
    L_out += at_least_one_bounce_radiance(r, isect);

  return L_out;
}

For indirect lighting, we recursively sample from the prior bounce’s location and direction. To make a proper estimator for the indirect lighting L, we have to first evaluate the BSDF’s reflectance f along with the a cosine term using the direction given from our sample wi, and covert those direction to world coordinates. We also scale by the pdf of the sample_f and the cpdf which is our normalization for russian roullete. The russian roullete randomly terminates the number of bounces.

Vector3D PathTracer::at_least_one_bounce_radiance(const Ray &r,
                                                  const Intersection &isect) {
  Matrix3x3 o2w;
  make_coord_space(o2w, isect.n);
  Matrix3x3 w2o = o2w.T();

  Vector3D hit_p = r.o + r.d * isect.t;
  Vector3D w_out = w2o * (-r.d);
  Vector3D L_out;

  L_out = one_bounce_radiance(r, isect);

  double cpdf = 0.8;

  Vector3D wi;
  double pdf;

  if (coin_flip(cpdf) && r.depth > 1) {
      Vector3D f = isect.bsdf->sample_f(w_out, &wi, &pdf);
      float costheta = wi.z;
      Ray ray_after_bounce = Ray(hit_p, o2w * wi);
      ray_after_bounce.depth = r.depth - 1;

      ray_after_bounce.min_t = EPS_F;

      Intersection isect_after_bounce;

      if (bvh->intersect(ray_after_bounce, &isect_after_bounce) && costheta > 0) {
          Vector3D L = at_least_one_bounce_radiance(ray_after_bounce, isect_after_bounce);
          L_out += L * f * costheta / pdf / cpdf;
      }
  }

  return L_out;

Results

Here are two results rendered at 1024 samples

The other results are the CBbunny.dae which is in the next sections

Light Bounce Progression

Here we show the progression of the first three light bounces individually for CBbunny.dae

0th Light Bounce

1st Light Bounce (Direct Lighting)

2nd Light Bounce (Inderect Lighting)

Indirect lighting has a lot less light since it only comes frrom diffuse sources, not direct illumination.

Increasing Depth

Now, we can show how increasing the depth of bounces allows for more and more realistic lighting. I altered russian roulette to have a 100% chance for these renders continuing.

Depth = 0

Depth = 1

Depth = 2

Depth = 3

Depth = 100

Notice that the result plateaus as depth increases. This is because the materials in the scene are all lambertian, making the energy of the light very weak so by the 100th bounce there is little light left to illuminate the scene.

We also increased the russian roullete progressively so that we are more likely to reach higher numbers of bounces. i.e cpdf = 0.99 for 100 bounces.

Sample-per-Pixel Rates

Here we vary samples to show how noise reduces. We use 4 light rays.

1 Sample

2 Samples

4 Samples

8 Samples

16 Samples

32 Samples

64 Samples

1024 Samples

As we expect, increasing samples allows for cleaner images.

Part 5

With indirect lighting, our renders now look more realistic. However, they are still noisy for lower sample rates. We want high sample rates for locaticons that need it (like shadows) and low sample rates for locations that converge quickly. This is what adaptive sampling does.

We want to stop our sampling when we have a value that is within the 95% confidence interval around the mean of the samples $\mu$.

Formally, we want the value $v$ of the pixel to be in between

$$\mu - I < v < \mu + I$$

Where $$I = 1.96 \cdot \frac{\sigma}{\sqrt{n}}$$

where $\sigma$ is the standard deviation of the points sampled and $n$ is the number of samples.

We then terminate the sampling if
$$I \leq maxTolerance \cdot \mu$$

Where $maxTolerance$ is a choice, usually arround 0.05.

We implement this in raytrace_pixel when looping over the samples. We check every samplesPerBatch, to not slow down the render too much. s1 is a sum of the illuminations and s2 is a sum of the illuminations squared.

for (int i = 1; i <= num_samples; i++, num_sampled++) {
      if (i % samplesPerBatch == 0 && i > 1) {
          float mu = s1 / (float)i;
          float var = (s2 - ((s1 * s1) / i)) / ((float)i - 1);
          float I = 1.96 * sqrt(var / (float)i);
          if (I <= maxTolerance * mu){
              break;
          }
      }

    auto sample_point = origin + gridSampler->get_sample();
    auto ray = camera->generate_ray(sample_point.x / sampleBuffer.w, sample_point.y / sampleBuffer.h);
    ray.depth = max_ray_depth;

    Vector3D radiance_est = est_radiance_global_illumination(ray);
    float illum_est = radiance_est.illum();

    s1 += illum_est;
    s2 += illum_est * illum_est;

      radiance += radiance_est;
  }

We also need to make sure to normalize the final of illumination by the number of samples actually sampled, not the total number of samples possible!

sampleBuffer.update_pixel(radiance * (1.0 / num_sampled), x, y);
sampleCountBuffer[x + y * sampleBuffer.w] = num_sampled;

Results

These renders were rendered with 2048 samples, 1 light ray, 5 max depth, and maxtolerance of 0.05.

Bunny Sample Rate

Bunny Rendered Result

Spheres Sample Rate

Spheres Rendered Result

Notice in both the most samples are where the indirect lighting dominates. That being shadows and the ceiling.

Collaboration

Tianchen coded parts 1 thru 3.3 and Riley coded parts 3.4 thru 5. We worked together on debugging individual parts.

This was the most interesting project so far since it does so many complex things that the results are often surprising.

Link:
https://cal-cs184-student.github.io/project-webpages-sp23-CardiacMangoes/proj3-1/index.html