Multicore Bundle Adjustment - Cse.unr.edu

1y ago
3 Views
2 Downloads
1.00 MB
15 Pages
Last View : 16d ago
Last Download : 2m ago
Upload by : Oscar Steel
Transcription

Multicore Bundle AdjustmentC. Wu, S. Agarwal, B. Curless, and S. SeitzUniversity of Washington * GooglePresented by Rory PierceCS791v Parallel ProgrammingInstructors Dr. Fred Harris and Lee BarfordOctober 4, 2011

Outline Bundle Adjustment Overview Approach Hardware Experiments Results

Bundle Adjustment Overview What is it? Reconstructing 3D structure and camera pose andparameters from a series of monocular or stereo images Why use it? Reconstruct object geometries Localize tourist photos Improve visual odometry estimates in visual SLAM

Bundle Adjustment Overview "Bundles" of light rays emanating from 3D objects convergeon the optical center of each camera, forming an image in a2D camera plane (projection). Estimated 3D points are projected onto 2D imageswhereupon they are compared with observed image points(re-projection error).

Bundle Adjustment Overview Problem formulation Non-linear least squares optimization Error is the L2 norm Implementation Levenberg-Marquardt (LM) Regularized linear approximations Solving and factoring large, sparse systems of linearequations

Approach Combining LM with Conjugate Gradients (CG) Implicit-Hessian Implicit-Schur (converges faster) Key point Suit parallelization Major computational expenseSerial time distribution

Approach Optimize memory Block Compressed Sparse Row (BCRS) matrices Optimize processing SSE on CPU with memory aligned parameters GPU threading scaled on number of cameras GPU problem as 1 thread per parameter, 1-half-warpper camera Use row access pattern Texture memory as cache when coalesced memory notapplicable

Approach Reduce dependency between threads Explicit computation and storage of gather indexes Reduce memory usage On-the-fly computations Utilize single-precision types Preserve accuracy through parameter distributionnormalization Revelation Matrix-free algorithms led to trade-off between time andmemory in CPU world Observed savings in both in GPU world

Hardware 2 Quad-core Intel Xeon E5520s (2.27 GHz with 2xhyperthreading) nVidia Tesla C1060 4 GB memory*64-bit Linux OS System

Experiments Benchmark Bundle Adjustment in the Large (BAL) [Agarwal et al.] Double-precision computation only 50 LM iterations Handicap Subject algorithm Single-precision for 100 LM iterations

Results GPU 10x-30x speedup Multi-threaded CPU 5x-10x speedup Consistent across Varying problem sizes Varying sparsity Ratios between GPU/CPU Highest on medium-sized problems Cost of accessing GPU memory increases due tolimited size of global texture cache on GPU

Results - Timings Units: Seconds CPU is SSE-optimized

Results - Large-scale problems GPU speed-up of 10x-30x Multi-threaded CPU solver 1/3 speed of GPU implementation

Results - Small-scale problems Consistent speedup via multi-core implementations Multi-core implementations do not reach same precisioncompared to BAL's double precision implementation

Final SlideQuestions?

Multicore Bundle Adjustment C. Wu, S. Agarwal, B. Curless, and S. Seitz University of Washington * Google Presented by Rory Pierce CS791v Parallel Programming Instructors Dr. Fred Harris and Lee Barford October 4, 2011. Outline Bundle Adjustment Overview . Large-scale problems GPU speed-up of 10x-30x Multi-threaded CPU solver 1/3 speed of .

Related Documents:

92 vipul sharma it 93 rishabh jain cse 94 manik arora cse 95 nishant bhardwaj cse . 96 rajit shrivastava it 97 shivansh gaur cse 98 harsh singh cse 99 shreyanshi raj cse 100 rahul bedi cse 101 pallavi anand cse 102 divya cse 103 nihal raj it 104 kanak

cse-148 kuriakose jijo george n t george cse-149 kusum joshi ramesh chandra joshi cse-150 m mithun bose n k mohandasan cse-151 madhuri yadav rajbir yadav cse-152 malini shukla r s sharma cse-153 manisha khattar sunil kumar khattar cse-154 m

University of Nevada Reno (UNR) Katia Albright UNR Handshake 775-682-7114 kalbright@unr.edu . NSHE Apprenticeship Navigator Workforce Development and Community Colleges 775-784-3410 colson@nshe.nevada.edu UNR Career Studio Katia Albright, Internship Development and Employer Relations Nevada Career Studio, the

Internships - Are You Involved? 7,500 Companies -Only 1,097 Interns Intern Programs TMCC Clinical UNR Pack Grant UNR Academic Credit UNR Career Studio WCSD WNC For Contacts go to EDAWN Employer Workforce Guide

This guide provides graduate students of the University of Nevada, Reno (UNR) with information on the structure of the MA and PhD programs in anthropology. It is provided to assist students in completing their work as expeditiously and as meaningfully as possible while complying with general UNR graduate school regulations. To

to adjust the sequence of packets by using the multicore NPU. Specifically, the contributions of this paper are threefold. † First, a multicore NPU-based stream reassembly architecture is introduced. To the best of our knowl-edge, this is the first work on employing multicore NPU-based stream reassembly technology specifi-cally for NIDS .

- A performance study of AMG on a large multicore cluster with 4-socket, 16-core nodes using MPI, OpenMP, and hybrid programming; - Scheduling strategies for highly asynchronous codes on multicore platforms; - A MultiCore SUPport (MCSup) library that provides efficient support for mapping an OpenMP program onto the underlying architecture;

Fiction Excerpt 1: The Adventures of Tom Sawyer (retold with excerpts from the novel by Mark Twain) Saturday morning was come, and all the summer world was bright and fresh, and brimming with life. There was a song in every heart; and if the heart was young the music issued at the lips. There was cheer in every face and a spring in every step. The locust trees were in bloom and the fragrance .