Class Project CS766

Multi-View 3D Geometry Reconstruction: Exploiting Massive Parallelism

Chaman Singh Verma

bascomBascom Hall generated by 150 images (Camera: Canon SX 100) using Bundler and PMVS2


Contents:

  1. Introduction

  2. Parallel Paradigms

    1. Fine Grain v/s Coarse Grain Parallelism.

    2. Centralized v/s Decentralized Scheduling.

  3. Exploring Parallelism

    1. Feature Detection.

    2. Feature Matching.

    3. Bundle Adjustment.

    4. PMVS.

    5. Surface Reconstruction.

  4. Results

  5. Future Work

  6. References


Results:

DataSet

#Images

Image Size
Middlebury: Temple
 359
 640x480
Middlebury: Dino
 363
 640x480
Ponce: GreenDragon
 24
 3104x2072
Pone: Armor
 48
 3504x2336
UW: BascomHall
 150
 1645x970
MadisonCapitol
 246
 1649x970
Chazan: Ganesh
 100
Chazan: Buddha
 100

capitol

Capitol Dataset: Output Point Cloud from Bundler + PMVS2


armor1
 armor2
 armor3
 armor4
 armor5

Armor Dataset: Output Point Cloud from Bundler + PMVS2


dragon1
 dragon2
 dragon3
 drago4n
 dragon5

Green Dragaon DataSet: Output from Point Cloud from Bundler + PMVS2

Middlebury Benchmark Dataset: (A) Temple (B) Dino

Execution time

DataSet

Feature Detection

( Intel Quad-Core: Hulk)

Feature Matching

(Condor 48 Nodes)

Bundle Adjustment

(Sequential )

PMVS2

(Intel QuadCore:Hulk)

Surface Reconstruction

(Sequential)
Temple
 125s
 398 s
 64 m

Dino
 96s
 68 s
 92 m

Dragon
 440s
 91s
 2m24s
 real 7m: user 23m
Armor
 983s
 671s
 3m30s

BascomHall
 792s
 2418 s
 72m
 real 186m: user 662m
Capitol
 1340s
 3294s
 242m

Ganesh




Buddha





Feature Detection:



Dataset

Min #Features

Max #Features

Mean #Features

Total Features
Temple
 445
 1144
 869
 314426
Dino
 66
 248
 156
 58654
Dragon
 6928
 9546
 8194
 198050
Armor
 11919
 55417
 27764
 1408775
BascomHall
 3673
 69907
 15120
 2317129
Capitol
 144
 28597
 13700
 3103375
IndianGoddess




  

DataSet/#Threads

1

2

4

6
Temple
 12/1.00
 6.0/2.00
 3.0/4.00
 2.20/5.45
Dino
 538/1.000
 271/1.985
 138/3.895
 95/5.663
Dragon
 831/1.000
 432/1.923
 237/3.506
 173/4.803
BascomHall
 4492/1.000
 2270/1.978
 1163/3.861
 776/5.787
Capitol
 7757/1.000
 3902/1.987
 1967/3.943
 1366/5.678
IndianGoddess




                                         SIFT Feature detection: Absolute time and Speedup on Multicore Machines on various dataset.

The results shown in the table shows that performance degrades as the number of threads increases supporting the intuitions that multicore machines do not scale well.

Feature Matching:

 

DataSet/#Processors

1

4

7
Temple
 113m 10s
 29m
 16m 41s
Dino
 5m 25s
4m 15s
 2m 38s
Dragon
 5m 25s
 1m 39s
 1m 7s
Armor
 105m 54s
 26m
 16m 5s
BascomHall
 6h 46m
 1h 45m
 62m 31s
Capitol
 14h 58m
 3h 50m
 137m3s
Ganesh


                                                                       load

Future Work: