Rendering
In Fall 2022, I took the COSC 287 Rendering Algorithm course at Dartmouth, taught by Wojciech Jarosz. This was my first foray into the field, and I plan to continue exploring it in the future. For my final project, I implemented volumetric photon mapping to create an Aurora effect.
Entire project report is shown below.
During the summer break of 2023, I developed a renderer using Metal Shader Language for Mac GPUs. The first version was released on July 18th and supports multiple ray-tracing methods and 7 built-in scenes. It also features a simple Swift-UI that allows for interactive changes to the settings.
Computer Graphics
In Winter 2023, I took the COSC 177 Computer Graphics course taught by Bo Zhu. This introductory-level course used OpenGL as the base code. For our final project, my group member and I used Ray Marching to complete the rendering, as shown below.
Computational Photography
In Winter 2023, I took the COSC 273 Computational Photography course taught by Adithya Pediredla. We covered topics such as pinhole cameras, HDR, light fields, and photometric stereo. For my final project, I transformed a video into a long exposure image and experimented with using a LYTRO light field camera to create a fake motion blur. I was the winner of the final project competition!
Deep Learning Projects
For my undergraduate thesis, I designed a network to compare image pairs based on the Mixture-of-Experts (MMoE) network. The goal was to detect data drift, which is important when deploying machine learning networks. I worked on this project as an intern at Volkswagen from October 2021 to May 2022. The network was designed to measure differences between two images in multiple fields, such as weather and lighting. The network design poster and presentation recording are listed below. (The Chinese version was translated by Google).
I also interned in 58.com AI-Lab in summer 2021. I mainly focused on the recommendation system. I did several experiments to match salesperson and customer. The model is based on DSSM (two-tower model). I did several iteration to get a better recall result (concerned about how many recommendations are provided among all the relevant recommendations). The model modification is shown as figure below.
