Dissertation Talk: Algorithm/Hardware Techniques for Nonlinear Least-Squares Problems in Computer Vision

Event details Date: Friday, July 24, 2026 Time: 2:00 PM to 2:00 PM Location: Gateway 1420. Zoom: https://berkeley.zoom.us/j/8742477341 Type: Lecture / Workshop Audience: Faculty,Students About this event Nonlinear least-squares (NLS) problems have been a staple in computer vision for a long time, from bundle adjustment in the 1950s to 3D Gaussian Splatting (3DGS) in the present day. Accelerating these applications, both in algorithm and hardware, can lead to significant improvements in modern computer vision pipelines. The key to this effort is in understanding the special structures and data representations in these problems. In this talk I will present three of my works in this area. First, I will showcase MAVERIC, a custom silicon tapeout for real-time simultaneous localization and mapping (SLAM). Next, I will discuss RA-ISAM2, a resource-aware incremental factorization algorithm designed for multicore, multi-accelerator systems-on-chips (SoCs). Finally, I will present 3DGS^2-TR, a scalable second-order algorithm for 3DGS scene training on commercial GPUs. Official event details: https://events.berkeley.edu/eecs/event/324550-dissertation-talk-algorithmhardware-techniques-for