A Calibrated Defocus Simulator for Research in Passive Ranging
CitationZhou, Dylan. 2022. A Calibrated Defocus Simulator for Research in Passive Ranging. Bachelor's thesis, Harvard College.
AbstractDepth from defocus is a prominent area of computer vision research, yet unlike many other areas of the field, there is no dedicated dataset or general library for training and testing depth from defocus applications. As a result, researchers have had to manually capture data, which takes a tremendous amount of patience and precision, or they have resorted to approximations of real-world data using self-designed simulation pipelines---this simulated data is often simplistic or lacking in physical accuracy, which can adversely affect the performance of depth from defocus algorithms in deployment.
This thesis proposes a new pipeline for generating simulated data for depth from defocus, one that uses physics-based rendering and has controllable defocus that can be calibrated to the optical dimensions of any physical system. This new pipeline promises to deliver all the efficiency of synthetic data generation, to generate data that is faithful to real-world physics, and to be generalizable across different depth from defocus applications while also maintaining flexibility in its specificity for different scene and object types.
Here, I explore the power of physics-based rendering for generating realistically defocused images and I present experiments and demonstrations to showcase its functionalities, capabilities, and limitations when it comes to generating synthetic data for depth from defocus research.
Citable link to this pagehttps://nrs.harvard.edu/URN-3:HUL.INSTREPOS:37371763
- FAS Theses and Dissertations