Graph Based Processing of Big Images
summary
Rewrites common image tasks (TV denoising, Poisson blending, spectral segmentation) as a single graph Laplacian linear system so one solver can be reused; implements reweighted L2 to approximate L1 and runs GPU-accelerated conjugate gradient or CMG for grid graphs. Prototypes combine denoising+segmentation and recover shadows/camouflage with better detail; scales strategy: GPU for <10 GB, in-memory bulk-synchronous for 10 GB–1 TB, asynchronous disk for >1 TB. Targets 2 GB under 5 minutes and demonstrates practical gains versus naïve map‑partition processing.
AI-written summary · about
chapters (14)
- 0:00Intro
- 2:23Commercial Sensors Today
- 4:57Signal Processing 101
- 6:58Compressive Sensing Remark Assumes sparsity and min separation of signals Reduces the number of samples required to reconstruct
- 8:32Mumford Shah, 1989
- 9:35Rudin Osher Fatemi, 1992
- 16:19Solving Linear Systems: Matrix Multiply to Electrical Circuits
- 17:30Graph Based Approach
- 25:07Graph Based Processing 1. Take advantage of "optimal" linear solvers. 2. Leverage on new graph technologies developed by the machine learning and Big Data community
- 25:43Graph Technologies Technology binned by problem sie
- 29:32Applications Graph Based Image processing allows the mix and match of various processing techniques Solver allows control of the degree which a technique is applied
- 30:13Spectral Segmentation
- 32:05Camouflage Detection
- 33:54Shadow Enhancement
related talks
24mTiling & Parallel Processing of Large Images
44mScalable Algorithmic Primitives for Data Science
17mBGL: GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
15mEfficient GPU-Centric Evolving Graph Processing at Scale
12mEfficient Workflow Serving for Diffusion Models with Many Adapters
34mDeep Learning at Scale
15mMGG: Accelerating Graph Neural Networks with Fine-Grained Intra-Kernel Communication...
19mOptimizations and abstractions for sparse machine learning