An optix GPU based path tracer. This is the renderer built and used in the following 2 projects:

• Li, Z., Shafiei, M., Ramamoorthi, R., Sunkavalli, K., & Chandraker, M. (2020). Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 2475-2484).
• Li, Z.*, Yeh, Y. Y.*, & Chandraker, M. (2020). Through the Looking Glass: Neural 3D Reconstruction of Transparent Shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 1262-1271).

Please contact Zhengqin Li about any questions with regard to the renderer. Please consider citing the 2 papers if you find the renderer useful in your own project.

The requirements are almost the same as optix_advanced_samples. We also need opencv 2 for image processing. ccmake is needed to compile the code. Please check INTALL-LINUX.txt and INSTALL-WIN.txt for details. After compiling, the executable file will be ./src/bin/optixRenderer

To run the code, use the following command

./src/bin/optixRenderer -f [shape_file] -o [output_file] -c [camera_file] -m [mode] --gpuIds [gpu_ids]

• shape_file: The xml file used to define the scene. See Shape file
• output_file: The name of the rendered image. For ldr image, we support formats jpg, bmp, png. For hdr image, currently we support .rgbe image only.
• camera_file: The file containing all the camera positions. We support rendering multiple images from different view points for a single scene. The format camera file will be

  N           # Number of view points
  a_0 a_1 a_2 # The origin of the camera 1
  b_0 b_1 b_2 # The lookAt of the camera 1
  c_0 c_1 c_2 # The up vector of camera 1
  .....
  

  The camera position can be also defined in shape file (.xml). But if camera file is given, the camera position in the shape file will be neglected.
• mode: The type of output of renderer
  • 0: Image (ldr or hdr)
  • 1: baseColor (.png)
  • 2: normal (.png)
  • 3: roughness (.png)
  • 4: mask (.png, If a ray from the pixel can hit an object, then the value of the pixel will be 255. If the ray hit an area light source or nothing, then the value will be 0)
  • 5: depth (.dat)
  • 6: metallic (.png)
• gpuIds: The ids of devices used for rendering. For example --gpuIds 0 1 2, then 3 gpus will be used for rendering.

There are some other flags to control the program.

• --noiseLimit: If the variance of image is larger than this threshold in the first iteration of adaptive samplilng. This image will be abandoned directly.
• --vertexLimit: If the .obj or .ply file contains more vertices than the threshold, it will be abandoned.
• --intensityLimit: If the intensity of the rendered image is smaller than the threshold, it will be abandoned.
• --camStart: The start camera Id in the camera file.
• --camEnd: The end camera Id in the camera file.
• --forceOutput: Overwrite the existing file.
• --rrBeginLength: Start Roulette Roulette for light path with length longer than this threshold.
• --maxPathLength: The maximum length of the light path.

We use xml file to define the scene. The format will be very similar to mitsuba. The main elements of the shape file include integrator, sensor, material, shape and emitter. We will introduce how to define the 5 elements in details in the following. An example rendering file can be find in ./src/optixRenderer/cbox. A typical shape file will look like:

<?xml version="1.0" encoding="utf-8"?>

<scene version="0.5.0">
  <integrator .../>
  ...
  <sensor .../>
  ...
  <bsdf .../>
  ...
  <emitter ../>
  ...
  <shape ../>
  ...
</scene>

Currently we only support path tracer. So the only acceptable way to define the integrator will be

<integrator type="path"/>

We support three types of sensors, the panorama camera (envmap), the hemisphere camera (hemisphere) and the perspective camera (perspective). The sub-elements belong to sensor include

• fovAxis: The axis of field of view. The value can be x or y.
• fov: Field of view, the measurement is degree.
• transform: The extrinsic parameter of the camera
• sampler: Currently we support two types of sampler. independent will actually do stratified sampling. adaptive sampling will render two images with different random seeds, scale the two images so that the mean value of pixel will be 0.5 and then compute the variance. If the variance is larger than a threshold, we will average the two images and render a new images by doubling the number of samples. We repeat this process until the variance is smaller than the threshold or the number of samples is too large. Adaptive sampling is highly recommended when rendering dataset.
• film: We currently support two types of film. ldrfilm will render ldr image with gamma correction (2.2). hdrfilm will render hdr image (.rgbe). The height and width of the images are set here.

Following is an example of the xml file of a perspective sensor. The xml for envmap and hemisphere camera should be almost the same, except that fov and fovAxis is no longer useful.

<sensor type="perspective">
  <string name="fovAxis" value="x"/>
  <float name="fov" value="60.0"/>
  <transform name="toWorld">
    <lookAt origin="278, 273, -800" target="278, 273, -799" up="0, 1, 0"/>
  </transform>
  <sampler type="independent">
    <integer name="sampleCount" value="10000"/>
  </sampler>

  <film type="ldrfilm">
    <integer name="width" value="512"/>
    <integer name="height" value="512"/>
  </film>
</sensor>

The renderer currently supports three kinds of materials: diffuse, phong and microfacet

• diffuse: The parameters include reflectance and normal. The type of reflectance can be rgb or texture. The type of normal can only be texture. The normal map should be a bitmap.
• phong: The parameters include diffuseReflectance, specularReflectance, alpha and normal. The type of diffuseReflectance and specularReflectance can be rgb or texture. The type of alpha can be float or texture.
• microfacet: The parameters include albedo, normal, roughness, metallic and fresnel. The type of albedo can be rgb or texture. The type of roughness, metallic and fresnel can be texture and float.
• dielectric: The parameters include specularReflectance, specularTransmittance, normal, intIOR and extIOR. The type of specularReflectance and specularTransmittance can only be rgb. The type of intIOR and extIOR can only be float.
• conductor: The parameters include specularReflectance. The type of specularReflectance can only be rgb.

Following is an example of phong material. Notice that the path to the texture should be absolute path.

<bsdf id="Model#317-1_material_28" type="phong">
  <texture name="diffuseReflectance" type="bitmap">
    <string name="filename" value="/home/exx/Zhengqin/SceneMaterial/Dataset/texture/wood_5.jpg"/>
  </texture>
  <rgb name="specularReflectance" value="0.43922 0.43922 0.43922"/>
  <float name="alpha" value="40.00000"/>
</bsdf>

Currently we support .obj file and .ply file formats. .mtl file is not supported so the material can only be defined in .xml file. However, we support using usemtl ... in .obj file so that you can have different materials for a single .obj file. Following is an example of a shape with two different materials. In .obj file, the material is defined using

...
usemtl Model#317-1_material_28
...
usemtl Model#317-1_material_29
...

The corresponding .xml file should be

<shape id="Model#317-1_object" type="obj">
  <string name="filename" value="/home/exx/Zhengqin/SceneMaterial/Dataset/house_obj_render/train/Shape__0004d52d1aeeb8ae6de39d6bd993e992/Model#317-1_object.obj"/>
  <bsdf id="Model#317-1_material_27" type="phong">
    <rgb name="diffuseReflectance" value="0.27843 0.27843 0.27843"/>
    <rgb name="specularReflectance" value="1.00000 1.00000 1.00000"/>
    <float name="alpha" value="300.00000"/>
  </bsdf>
  <bsdf id="Model#317-1_material_28" type="phong">
    <texture name="diffuseReflectance" type="bitmap">
      <string name="filename" value="/home/exx/Zhengqin/SceneMaterial/Dataset/texture/wood_5.jpg"/>
    </texture>
    <rgb name="specularReflectance" value="0.43922 0.43922 0.43922"/>
    <float name="alpha" value="40.00000"/>
  </bsdf>
</shape>

Notice that the id of brdf should be consistent with the material defined by usemtl commmand. We also support scale, translate and rotate an object. The format is exactly the same as mitsuba.

We support three kinds of emitter, area light (area), point light (point), flash light (flash) and environment map (envmap).

• area: The area light can be considered as an sub-elements of shape. It has one parameter radiance. The type of radiance is rgb
• point and flash: The point light has two parameters. The radiance controls the light intensity while position controls its position in the scene. The type of position is point. The flash light only has parameter radiance. The position will be the same as origin of camera. That will be useful when rendering with camera file.
• envmap: The image used for environmental map must be of format .hdr. The scale parameter can be used to control the intensity of the environmental map.

Followings are examples of area light and envmap respectively.

<shape type="obj">
  <string name="filename" value="/home/exx/Zhengqin/SceneMaterial/Code/optix_advanced_samples/src/optixSUNCG/cbox/meshes/cbox_luminaire.obj"/>
  <emitter type="area">
    <rgb name="radiance" value="20.0 20.0 20.0"/>
  </emitter>
</shape>

<emitter type = "envmap">
  <string name="filename" value="/home/exx/Zhengqin/SceneMaterial/Code/optix_advanced_samples/src/optixSUNCG/cbox/envmap.hdr"/>
</emitter>

There are three .xml files in src/optixRenderer/cbox directory, which can be directly used for testing.

• Bidirectional path tracing should be added, if we have time in the future.