Please note: This PhD seminar will take place online.
Xiaochun Tong, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Toshiya Hachisuka
Differentiable rendering has enabled gradient-based optimization of complex image synthesis processes, with applications ranging from inverse graphics to computational optical design. While significant progress has been made in differentiating continuous parameters, even in the presence of scene discontinuities, many practical design problems also involve discrete structural choices that cannot be handled by existing differentiable rendering techniques. Compound lens design is one such problem, where both continuous optical parameters and discrete lens configurations must be considered.
This work presents a method for jointly optimizing continuous and discrete parameters in compound lens design to improve optical sharpness, speed, or both. Existing approaches primarily rely on gradient-based optimization of continuous parameters, such as the curvature of individual lens elements, while changes to lens topology typically require substantial expert intervention. In contrast, this work also optimizes discrete parameters, including the number and type of lens elements.
The proposed method combines gradient-based optimization with a Markov chain Monte Carlo algorithm. Continuous parameters are optimized with gradient descent while discrete structures are explored with MCMC mutations. Novel mutations strategies are proposed to aid global exploration in a transdimensional setting. Experiments across a variety of lens design tasks demonstrate that the method effectively explores a broader design space, producing better lens configurations than previous approaches and improving the achievable trade-offs between optical speed and sharpness.