updates to renderer doc text

Summary: Pull Request resolved: #3

Reviewed By: gkioxari

Differential Revision: D19523239

Pulled By: nikhilaravi

fbshipit-source-id: 46b3d52cdabbddb4fecba6038f6ca23a0b736571

Author: nikhilaravi

docs/figs/subset_batch_size_128.png

@@ -1,12 +1,18 @@
 # Differentiable Rendering
 
-Differentiable rendering is an exciting research area but it is not yet clear what the best practices are. We extensively researched existing codebases and found that:
+Differentiable rendering is a relatively new and exciting research area in computer vision, bridging the gap between 2D and 3D by allowing 2D image pixels to be related back to 3D properties of a scene.
+
+For example, by rendering an image from a 3D shape predicted by a neural network, it is possible to compute a 2D loss with a target image. Inverting the rendering step means we can relate the 2D loss from the pixels back to the 3D properties of the shape such as the positions of mesh vertices, enabling 3D shapes to be learnt without any explicit 3D supervision.
+
+We extensively researched existing codebases for differentiable rendering and found that:
 - the rendering pipeline is complex with more than 7 separate components which need to interoperate and be differentiable
-- popular existing approaches [[1](#1), [2](#2)] are based on the same core implementation which bundles many of the key components into large CUDA kernels which require significant expertise to understand and has limited scope for extensions  
+- popular existing approaches [[1](#1), [2](#2)] are based on the same core implementation which bundles many of the key components into large CUDA kernels which require significant expertise to understand, and has limited scope for extensions  
 - existing methods either do not support batching or assume that meshes in a batch have the same number of vertices and faces
 - existing projects only provide CUDA implementations so they cannot be used without GPUs
 
-In order to experiment with different approaches we wanted a modular implementation that is easy to use and extend and supports [heterogeneous batching](batching.md).  Taking inspiration from existing work in this area [[1](#1), [2](#2)] we have created a new modular, differentiable renderer with **parallel implementations in PyTorch, C++ and CUDA**.
+In order to experiment with different approaches, we wanted a modular implementation that is easy to use and extend, and supports [heterogeneous batching](batching.md).
+
+Taking inspiration from existing work [[1](#1), [2](#2)], we have created a new, modular, differentiable renderer with **parallel implementations in PyTorch, C++ and CUDA**, as well as comprehensive documentation and tests, with the aim of helping to further research in this field.
 
 Our implementation decouples the rasterization and shading steps of rendering. The core rasterization step (based on [[2]](#2)) returns several intermediate variables and has an optimized implementation in CUDA. The rest of the pipeline is implemented purely in PyTorch, and is designed to be customized and extended. With this approach, the PyTorch3d differentiable renderer can be imported as a library.
 
@@ -15,32 +21,39 @@ Our implementation decouples the rasterization and shading steps of rendering. T
 To learn about more the implementation and start using the renderer refer to [renderer_getting_started.md](renderer_getting_started.md), which also contains the [architecture overview](../figs/architecture_overview.png) and [coordinate transformation conventions](../figs/transformations_overview.png).
 
 
-##<u>Key features</u>
+## <u>Key features</u>
 
 ### 1. CUDA support for fast rasterization of large meshes
 
-We implemented modular CUDA kernels for the forward and backward pass of rasterization, adaptating a traditional graphics approach known as "coarse to fine" rasterization.
+We implemented modular CUDA kernels for the forward and backward pass of rasterization, adaptating a traditional graphics approach known as "coarse-to-fine" rasterization.
 
-First, the image is divided into a coarse grid and mesh faces are allocated to the grid cell in which they occur. This is followed by a refinement step which does pixel wise rasterization of the reduced subset of faces per grid cell. The grid size is a parameter which can be varied.
+First, the image is divided into a coarse grid and mesh faces are allocated to the grid cell in which they occur. This is followed by a refinement step which does pixel wise rasterization of the reduced subset of faces per grid cell. The grid cell size is a parameter which can be varied (`bin_size`).
 
-We additionally introduce a parameter `faces_per_pixel` which allows users to specify the top K faces which should be returned per pixel in the image (as opposed to traditional rasterization which returns only the index of the closest face in the mesh per pixel). The top K face properties can then be aggregated using different methods (such as the sigmoid/softmax approach proposed by Li et at in for SoftRasterizer [[2]](#2)).
+We additionally introduce a parameter `faces_per_pixel` which allows users to specify the top K faces which should be returned per pixel in the image (as opposed to traditional rasterization which returns only the index of the closest face in the mesh per pixel). The top K face properties can then be aggregated using different methods (such as the sigmoid/softmax approach proposed by Li et at in SoftRasterizer [[2]](#2)).
 
-We compared with the SoftRasterizer, to measure the effect of both these design changes on the speed of rasterizing a set of meshes of different sizes from ShapeNetV1 core. We rasterize one mesh in each batch to produce images of different sizes and measure the speed of the forward and backward passes.
+We compared PyTorch3d with SoftRasterizer to measure the effect of both these design changes on the speed of rasterization. We selected a set of meshes of different sizes from ShapeNetV1 core, and rasterized one mesh in each batch to produce images of different sizes. We report the speed of the forward and backward passes.
 
 **Fig 1: PyTorch3d Naive vs Coarse-to-fine**
 
-This figure shows how the coarse to fine strategy for rasterization results in significant speed up compared to naive rasterization. This is especially clear in large images.
+This figure shows how the coarse-to-fine strategy for rasterization results in significant speed up compared to naive rasterization for large image size and large mesh sizes.
 
 <img src="../figs/p3d_naive_vs_coarse.png" width="1000">
 
 
+For small mesh and image sizes, the naive approach is slightly faster. We advise that you understand the data you are using and choose the rasterization setting which suits your performance requirements. It is easy to switch between the naive and coarse-to-fine options by adjusting the `bin_size` value when initializing the [rasterization settings](https://github.com/facebookresearch/pytorch3d/blob/master/pytorch3d/renderer/mesh/rasterizer.py#L26).
+
+Setting `bin_size = 0` will enable naive rasterization. If `bin_size > 0`, the coarse-to-fine approach is used. The default is `bin_size = None` in which case we set the bin size based on [heuristics](https://github.com/facebookresearch/pytorch3d/blob/master/pytorch3d/renderer/mesh/rasterize_meshes.py#L92).
+
 **Fig 2: PyTorch3d Coarse-to-fine vs SoftRasterizer**
 
-This figure shows the speed up in the full forward and backward pass enabled by the combination of coarse-to-fine approach and caching the faces rasterized per pixel returned from the forward pass. In the SoftRasterizer implementation, in both the forward and backward pass, there is a loop over every single face in the mesh for every pixel in the image. Therefore, the time for the full forward plus backward pass is ~2x the time for the forward pass.  
+This figure shows the effect of the _combination_ of coarse-to-fine rasterization and caching the faces rasterized per pixel returned from the forward pass. For large meshes and image sizes, we again observe that the PyTorch3d rasterizer is significantly faster.
+
+In the SoftRasterizer implementation, in both the forward and backward pass, there is a loop over every single face in the mesh for every pixel in the image. Therefore, the time for the full forward plus backward pass is ~2x the time for the forward pass. For small mesh and image sizes, the SoftRasterizer approach is slightly faster.
 
 <img src="../figs/p3d_vs_softras.png" width="1000">
 
 
+
 ### 2. Support for Heterogeneous Batches
 
 PyTorch3d supports efficient rendering of batches of meshes where each mesh has different numbers of vertices and faces. This is done without using padded inputs.
@@ -49,37 +62,28 @@ We again compare with SoftRasterizer which only supports batches of homogeneous
 
 We group meshes from Shapenet into bins based on the number of faces in the mesh, and sample to compose a batch. We then render images of fixed size and measure the speed of the forward and backward passes.
 
-We tested with a range of increasingly large bin sizes.
+We tested with a range of increasingly large meshes and bin sizes.
 
 **Fig 3: PyTorch3d heterogeneous batching compared with SoftRasterizer**
 
 <img src="../figs/fullset_batch_size_16.png" width="700"/>
 
-**Fig 3(a):** This shows that for large meshes and large bin width (i.e. more variation in mesh size in the batch) the heterogeneous batching approach in PyTorch3d is faster than either of the workarounds with SoftRasterizer.
-(settings: batch size = 16, mesh sizes in bins ranging from 500-350k faces, image size = 64)
-
-<img src="../figs/subset_batch_size_128.png" width="700"/>
-
-**Fig 3(b):** For larger batch sizes with smaller mesh sizes and bin sizes, PyTorch3d is still comparably fast with improved performance again in the cases of larger meshes and larger bin width.
-(settings: batch size in [64, 128] for mesh sizes in bins from 500-10k faces, image size = 128)
+This shows that for large meshes and large bin width (i.e. more variation in mesh size in the batch) the heterogeneous batching approach in PyTorch3d is faster than either of the workarounds with SoftRasterizer.
 
+(settings: batch size = 16, mesh sizes in bins ranging from 500-350k faces, image size = 64, faces per pixel = 100)
 
 ---
 
 **NOTE: CUDA Memory usage**
 
-The SoftRasterizer forward CUDA kernel only outputs one (N, H, W, 4) FloatTensor compared with the PyTorch3d rasterizer forward CUDA kernel which outputs 4 tensors:
-    - `pix_to_face`, LongTensor `(N, H, W, K)`  
-    - `zbuf`, FloatTensor `(N, H, W, K)`
-    - `dist`, FloatTensor `(N, H, W, K)`
-    - `bary_coords`, FloatTensor `(N, H, W, K, 3)`
+The SoftRasterizer forward CUDA kernel only outputs one `(N, H, W, 4)` FloatTensor compared with the PyTorch3d rasterizer forward CUDA kernel which outputs 4 tensors:
 
-The PyTorch3d backward pass returns gradients for:
-    - `zbuf`, FloatTensor `(N, H, W, K)`
-    - `dist`, FloatTensor `(N, H, W, K)`
-    - `bary_coords`, FloatTensor `(N, H, W, K, 3)`
+  - `pix_to_face`, LongTensor `(N, H, W, K)`  
+  - `zbuf`, FloatTensor `(N, H, W, K)`
+  - `dist`, FloatTensor `(N, H, W, K)`
+  - `bary_coords`, FloatTensor `(N, H, W, K, 3)`
 
-where **N** = batch size, **H/W** are image height/width, **K** is the faces per pixel.
+where **N** = batch size, **H/W** are image height/width, **K** is the faces per pixel. The PyTorch3d backward pass returns gradients for `zbuf`, `dist` and `bary_coords`.
 
 Returning intermediate variables from rasterization has an associated memory cost. We can calculate the theoretical lower bound on the memory usage for the forward and backward pass as follows:
 

docs/notes/renderer.md

@@ -41,7 +41,7 @@ Rendering requires transformations between several different coordinate frames:
 
 While we tried to emulate several aspects of OpenGL, the NDC coordinate system in PyTorch3d is **right-handed** compared with a **left-handed** NDC coordinate system in OpenGL (the projection matrix switches the handedness).
 
-In OpenGL, the camera at the origin is looking along -Z axis in camera space, but it is looking along +Z axis in NDC space.
+In OpenGL, the camera at the origin is looking along `-z` axis in camera space, but it is looking along the `+z` axis in NDC space.
 
 <img src="../figs/opengl_coordframes.png" width="200">