postPerspective

Optimizing Image Processing

Wednesday, 27 July, 9:00 am - 10:30 am, Anaheim Convention Center, Ballroom C
Session Chair: James Hays, Georgia Institute of Technology, Brown University

Image Perforation: Automatically Accelerating Image Pipelines by Intelligently Skipping Samples

Image perforation is a compilation method for accelerating image pipelines. It does this by combining an image pipeline with approximation rules, each of which causes sample computations to be skipped.

Liming Lou
University of Virginia, Shandong University

Paul Nguyen
University of Virginia

Jason Lawrence
University of Virginia

Connelly Barnes
University of Virginia

Automatically Scheduling Halide Image-Processing Pipelines

This algorithm for automatically generating schedules for Halide image-processing pipelines uses a simple, model-driven approach to implement key optimizations employed daily by expert Halide developers. It avoids costly auto-tuning and, in seconds, generates schedules competitive with those manually authored by expert Halide developers.

Ravi Mullapudi
Carnegie Mellon University

Andrew Adams
Google

Dillon Sharlet
Google

Jonathan Ragan-Kelley
Stanford University

Kayvon Fatahalian
Carnegie Mellon University

ProxImaL: Efficient Image Optimization Using Proximal Algorithms

ProxImaL is a domain-specific language and compiler for image-optimization problems that makes it easy to experiment with different problem formulations and algorithm choices. The compiler intelligently chooses the best way to translate a problem formulation and choice of optimization algorithm into an efficient solver implementation.

Felix Heide
Stanford University

Steven Diamond
Stanford University

Matthias Nießner
Stanford University

Jonathan Ragan-Kelley
Stanford University

Wolfgang Heidrich
King Abdullah University Of Science And Technology

Gordon Wetzstein
Stanford University

Rigel: Flexible Multi-Rate Image Processing Hardware

This paper presents a new high-level multi-rate architecture and compiler for image processing hardware that supports image pyramids, sparse computations, and space-time tradeoffs. It shows depth from stereo, Lucas-Kanade, and a Gaussian Pyramid running at 20-436 megapixels/second on two FPGA platforms.

James Hegarty
Stanford Univeristy

Zachary DeVito
Stanford University

Jonathan Ragan-Kelley
Stanford University

Pat Hanrahan
Stanford University

Ross Daly
Stanford University

Mark Horowitz
Stanford University