Presentation
Local-GS: An Order-Independent Gaussian Splatting Training Accelerator Exploiting Splat Locality
Description3D Gaussian Splatting recently emerged as the new SOTA approach for 3D representation and view synthesis. While Gaussian Splatting has demonstrated impressive training capability and rendering quality on desktop GPUs, achieving on-demand training on resource-constrained edge devices is still challenging. In this work, we identified the training bottleneck comes from a few perspectives including under-utilized redundant rendering threads and insufficient shared memory. To address these problems, we present Local-GS, a compact 3D Gaussian Splatting training accelerator utilizing order-independent rendering to break the depth-wise data dependency between overlapping Gaussians. We further incorporate a highly-parallel pixel intersection unit to reschedule thread workload and improve hardware utilization based on Gaussian locality. A set of compact unified training-rendering core is also designed to achieve efficient splat-level parallel rendering and gradient propagation. Local-GS is implemented and evaluated in 7nm technology with several real-world scenes, achieving training speed improvement of 26.9-53× across different scenarios compared to Jetson Xavier NX Mobile GPU.
Event Type
Research Manuscript
TimeWednesday, June 252:30pm - 2:45pm PDT
Location3000, Level 3
Similar Presentations