29
ProxImaL: Efficient Image Optimization using Proximal Algorithms Steven Diamond 1 Felix Heide 1,2 Wolfgang Heidrich 3,2 Gordon Wetzstein 1 2 University of British Columbia 3 KAUST 1 Stanford University www.proximal-lang.org Matthias Nießner 1 Jonathan Ragan-Kelley 1

ProxImaL | SIGGRAPH 2016

Embed Size (px)

Citation preview

Page 1: ProxImaL | SIGGRAPH 2016

ProxImaL: Efficient Image Optimization using Proximal Algorithms

Steven Diamond1Felix Heide1,2

Wolfgang Heidrich3,2 Gordon Wetzstein1

2University of British Columbia 3KAUST1Stanford University

www.proximal-lang.org

Matthias Nießner1 Jonathan Ragan-Kelley1

Page 2: ProxImaL | SIGGRAPH 2016

Low-Light Burst Imaging

Pelican Color Array

Interlaced HDR and RGB-IR

Light.co Array Camera Kinect ToF Depth

Imaging

Page 3: ProxImaL | SIGGRAPH 2016

Formal Optimization

Zoran and Weiss 2011 Levin et al. 2004Krishnan and Szeliski 2011

Krishnan and Fergus 2009Heide et al. 2015

Deconvolution Denoising Inpainting + Colorization Camera Image Processing

Schmidt et al. 2015 Chen et al. 2015

Page 4: ProxImaL | SIGGRAPH 2016

Demosaic Denoise

Bad Pixel Correction

Image Enhancing

Tone Mapping

Lens Correction

Black Level

Meteringerror error

errorerror

Formal Optimization

Image Processing Pipeline

Page 5: ProxImaL | SIGGRAPH 2016

Formal Optimization

Page 6: ProxImaL | SIGGRAPH 2016

Formal Optimization

Brooke et al. 1988 Grant and Boyd. 2014 Lofberg 2004

DSLs for convex optimization:

Page 7: ProxImaL | SIGGRAPH 2016

Formal Optimization

Brooke et al. 1988Grant and Boyd. 2014 Lofberg 2004

DSLs for convex optimization:

Infeasible for Imaging problems:• Millions of Variables• Large-Scale Operators

Page 8: ProxImaL | SIGGRAPH 2016

ProxImaL

Page 9: ProxImaL | SIGGRAPH 2016

ProxImaLAndroid HDR+First Frame

ProxImaL Code:

ProxImaLAndroid HDR+

Page 10: ProxImaL | SIGGRAPH 2016

Objective:

An example:

Proximal Code:

OriginalBlurredSubsampled

Page 11: ProxImaL | SIGGRAPH 2016

Translation “by Hand”:

Objective:

or: with either:

Page 12: ProxImaL | SIGGRAPH 2016

ADMM:

Objective:

or: with either:

100 sec 10 sec

Blurred

Blurred + Subsampled

Result

Ambiguous translations drastically affect solver performance !

Translation “by Hand”:

Page 13: ProxImaL | SIGGRAPH 2016

Sum of “proxable” functions:

General Problem Representation:

• . are “proxable” penalty functions with the proximal operator:

are linear transforms on the unknowns.• .

Proximal algorithms:

• ADMM [Boyd 2011]• Linearized ADMM [Boyd 2011]• PC [Chambolle and Pock 2011]• (HQS [Geman and Yang 1995])

Page 14: ProxImaL | SIGGRAPH 2016

Proximal Compiler:

Objective:

Page 15: ProxImaL | SIGGRAPH 2016

Algorithm Implementation:

Halide

Function Numpy [ms] Halide [ms]

sum_of_squares 246 42

dot product 97 16

subsample 356 73

grad 1188 95

conv 7791 121

warp 458 153

norm1 202 27

group_norm1 1037 68

FFT 23 9

Page 16: ProxImaL | SIGGRAPH 2016

Runtime of TV-Deconvolution:

Page 17: ProxImaL | SIGGRAPH 2016

Runtime of TV-Deconvolution:

Page 18: ProxImaL | SIGGRAPH 2016

Applications:

Demosaicking Interlaced HDR Low-Light Burst Imaging

Poisson Deconvolution

Phase Retrieval

Page 19: ProxImaL | SIGGRAPH 2016

Applications:

Demosaicking Interlaced HDR Low-Light Burst Imaging

Poisson Deconvolution

Phase Retrieval

Page 20: ProxImaL | SIGGRAPH 2016

ProxImaL

ProxImaL Code:

ProxImaLKrishnan and Fergus 2009

Page 21: ProxImaL | SIGGRAPH 2016

Applications:

Demosaicking Interlaced HDR Low-Light Burst Imaging

Poisson Deconvolution

Phase Retrieval

Page 22: ProxImaL | SIGGRAPH 2016

ProxImaL Code:

40 dB 34 dB

Page 23: ProxImaL | SIGGRAPH 2016

Applications:

Demosaicking Interlaced HDR Low-Light Burst Imaging

Poisson Deconvolution

Phase Retrieval

Page 24: ProxImaL | SIGGRAPH 2016

ProxImaL Code:

Page 25: ProxImaL | SIGGRAPH 2016

Applications:

Demosaicking Interlaced HDR Low-Light Burst Imaging

Poisson Deconvolution

Phase Retrieval

Page 26: ProxImaL | SIGGRAPH 2016

ProxImaL Code:

Page 27: ProxImaL | SIGGRAPH 2016

Applications:

Demosaicking Interlaced HDR Low-Light Burst Imaging

Poisson Deconvolution

Phase Retrieval

Page 28: ProxImaL | SIGGRAPH 2016

Applications:

Demosaicking Interlaced HDR Low-Light Burst Imaging

Poisson Deconvolution

Phase Retrieval

Please see paper !

Page 29: ProxImaL | SIGGRAPH 2016

ProxImaLwww.proximal-lang.org

Open Source !