May 5, differentiable rendering & unsupervised 3D reconstruction

Slides:

25 Data-driven shape modeling - deep learning, differentiable rendering.pdf Download 25 Data-driven shape modeling - deep learning, differentiable rendering.pdf

Panopto recording:

Learning check:

Read the paper "DIST: Rendering Deep Implicit Signed Distance Function with Differentiable Sphere Tracing Links to an external site.”, (CVPR 2020) and answer the questions below.

1. Describe the problem statement of the paper.
2. Explain Figure 2 in your own words.
3. A key challenge is the efficiency of sphere tracing, because evaluating the implicit function at each step requires a network evaluation.  A core idea of the paper is a number of techniques to make sphere tracing as efficient as possible. Describe each technique they propose in one or two sentences.
4. Their approach does not use an encoder-decoder network. Instead, given an input image they perform shape reconstruction by solving the optimization problem in Equation 5. Explain this equation in a few sentences. 
5. Formulate one question about the paper.

Comments and Q&A for learning check

https://docs.google.com/document/d/1qEvj-amWHFezmbgUExU36U_ekE52M4jFLfeQ2QIq7RU Links to an external site.