Overcoming Ceiling Performance: Using Complexity Filtering for Harder Inverse Graphics Benchmarks

Table of Links

Abstract and 1. Introduction

1. Background & Related Work

2. Method

   3.1 Sampling Small Mutations

   3.2 Policy

   3.3 Value Network & Search

   3.4 Architecture

3. Experiments

   4.1 Environments

   4.2 Baselines

   4.3 Ablations

4. Conclusion, Acknowledgments and Disclosure of Funding, and References

   \

Appendix

A. Mutation Algorithm

B. Context-Free Grammars

C. Sketch Simulation

D. Complexity Filtering

E. Tree Path Algorithm

F. Implementation Details

D Complexity Filtering

As mentioned in Section 4, while testing our method alongside baseline methods, we reached ceiling performance for all our methods. Ellis et al. [11] got around this by creating a “hard” test case by sampling more objects. For us, when we increased the number of objects to increase complexity, we saw that it increased the probability that a large object would be sampled and subtract from the whole scene, resulting in simpler scenes. This is shown by Figure 11(b), which is our training distribution. Even though we sample a large number of objects, the scenes don’t look visually interesting. When we studied the implementation details of Ellis et al. [11], we noticed that during random generation of expressions, they ensured that each shape did not change more that 60% or less than 10% of the pixels in the scene. Instead of modifying our tree sampling method, we instead chose to rejection sample based on the compressibility of the final rendered image.

\

:::info Authors:

(1) Shreyas Kapur, University of California, Berkeley (srkp@cs.berkeley.edu);

(2) Erik Jenner, University of California, Berkeley (jenner@cs.berkeley.edu);

(3) Stuart Russell, University of California, Berkeley (russell@cs.berkeley.edu).

:::

:::info This paper is available on arxiv under CC BY-SA 4.0 DEED license.

:::

\