Differentiable computer graphics for training and verification of machine perception

Differentiable computer graphics for training and verification of machine perception
Fredo Durand

We aim to rigorously characterize the sensitivity of vision-in-the-loop driving controllers in increasingly complex visual tasks. While rooftop lidar provides a spectacular amount of high-rate geometric data about environment, there are a number of tasks in an autonomous driving system where camera-based vision will inevitably play a dominant role: dealing with lane markings and road signs, dealing with water/snow and other inclement weather conditions that can confuse a lidar, and even dealing with construction (orange cones), police officers, and pedestrians/animals. Furthermore, vision sensors are often fused with depth returns from a laser and other sensors as a part of the vehicle and obstacle estimation algorithms.

As an initial study, we will investigate the performance of a simple perception algorithm for lane detection and a simple controller for lane following, given visual scenes which capture some of the diversity of urban driving conditions here in Boston including complex on-road traffic markings at intersections and worn visual features. As the project progresses, we will attempt to simulate more and more of the visual world -- up to and including difficult volumetric effects such as fog or snow and dynamic obstacles such as pedestrians and other vehicles.

This is a continuation of the project "Simulation and Verification for Vision-in-the-Loop Control" by Fredo Durand.

 

Publications:

  • T.-M. Li, M. Aittala, F. Durand, and J. Lehtinen, “Differentiable Monte Carlo Ray Tracing through Edge Sampling,” in ACM Trans. Graph. (Proc. SIGGRAPH Asia), 2018, vol. 37, pp. 222:1–222:11 [Online]. Available: https://doi.org/10.1145/3272127.3275109

  • T.-M. Li, M. Gharbi, A. Adams, F. Durand, and J. Ragan-Kelley, “Differentiable programming for image processing and deep learning in halide,” ACM Trans. Graph., vol. 37, no. 4, pp. 1–13, Jul. 2018 [Online]. Available: https://doi.org/10.1145/3197517.3201383. [Accessed: 16-Sep-2019]

  • M. Gharbi, J. Chen, J. Barron, S. Hasinoff, and F. Durand, “Deep Bilateral Learning for Real-Time Image Enhancement,” ACM Transactions on Graphics, vol. 36, no. 4, Jul. 2017 [Online]. Available: https://doi.org/10.1145/3072959.3073592, https://groups.csail.mit.edu/graphics/hdrnet/
  • M. Gharbi, T.-M. Li, M. Aittala, J. Lehtinen, and F. Durand, “Sample-based Monte Carlo denoising using a kernel-splatting network,” ACM Transactions on Graphics, vol. 38, no. 4, pp. 1–12, Jul. 2019 [Online]. Available: https://doi.org/10.1145/3306346.3322954

  • L. Anderson, T.-M. Li, J. Lehtinen, and F. Durand, “Aether: An Embedded Domain Specific Sampling Language for Monte Carlo Rendering,” ACM Transactions on Graphics, vol. 36, no. 4, Jul. 2017 [Online]. Available: https://doi.org/10.1145/3072959.3073704, https://people.csail.mit.edu/lukea/aether/
  • A. Zlateski, R. Jaroensri, P. Sharma, and F. Durand, “On the Importance of Label Quality for Semantic Segmentation,” in IEEE CVPR 2018, 2018 [Online]. Available: https://doi.org/10.1109/CVPR.2018.00160

     

 

Videos: