Final Project - Evan Rosson

VRIP Implementation

(Minus hole filling)

Evan Rosson

CSC 570

Winter 2007

Download
Project Goals
Methods
Documentation
Known Problems and Future Work
Results
Resources

Download

This program has been tested on Ubuntu Linux 6.10. You'll need OpenGL and GLUT. All other required libraries should be included, though you may have to play around with the makefiles for the libraries included in the source's /lib directory.

Source

Project Goals

Implement VRIP, without the hole filling, for educational purposes.

Methods

The program goes through five stages, all of which may be visualized:

Read each range image
Generate a surface from each range image
Create volume (d(x)/w(x)) from each surface
Merge volumes (derive D(x))
Generate merged surface (marching cubes)

Documentation

Command line usage:

./vrip [OPTIONS...] alignment.conf [data.ply...]

Ply files should be in ASCII format. (See the included Stanford Bunny data.)

Command Line Options
--res X Y Z	Volume resolution; X * Y * Z voxels. Default 10 * 10 * 10.
--cutoff VALUE	Weight cutoff value used when generating volume. Default 0.20. Used to prevent surfaces on opposite sides of the model from interacting
--zthresh VALUE	Z threshhold used when generating range surfaces. Default 0.005 Increase this value if your range surfaces are connected where no connection should appear.

Camera controls

a,d Move left/right

w,s Move up/down

z,x Move forward/backward

A,D Rotate along x-axis

W,S Rotate along y-axis

Z,X Rotate along z-axis

r Reset to starting camera angle/size

g Show next stage*

G Show previous stage*

o Show all input files

n Show only the next input file

h Display help

i Image alignment on/off

f Flatten image

q Scale image up (enlarge)

Q Scale image down (shrink)

c Toggle colors

l List alignment/display settings

* Stages: range image -> surface -> volume -> merged volume -> merged surface

Camera controls
a,d	Move left/right
w,s	Move up/down
z,x	Move forward/backward
A,D	Rotate along x-axis
W,S	Rotate along y-axis
Z,X	Rotate along z-axis
r	Reset to starting camera angle/size
g	Show next stage*
G	Show previous stage*
o	Show all input files
n	Show only the next input file
h	Display help
i	Image alignment on/off
f	Flatten image
q	Scale image up (enlarge)
Q	Scale image down (shrink)
c	Toggle colors
l	List alignment/display settings

Known Problems and Future Work

Ordered by importance relative to difficulty of implementation:

Volume construction has serious performance issues: every voxel is compared with every triangle when checking for an intersection. The result is ridiculously slow volume construction, even for low resolution volumes: a 25*25*25 volume takes around a half hour to render on my machine; a 50*50*50 volume requires an overnight run. Partitioning the triangles/voxels in space, or some clever sorting, would speed things up dramatically.
Add an option to construct the merged volume incrementally, instead of constructing separate volumes and merging them. This would prevent volumes from being visualized separately, but save a great deal of memory.
A more accurate formula for vertex weight should be used. This would not only create higher quality results, but this would allow us to use weight to automatically calculate cutoff values, as described in the original paper. (At the moment, cutoff values must be entered by the user, and getting a good cutoff value is largely guesswork.)
Triangle intersection/distance in volumes ought to be calculated more accurately: at the moment, some gross approximations are used that are hurting the quality of our results, even at higher resolutions.
Work with binary PLY files. (The Stanford Bunny's PLY files are in ASCII; most other Stanford data is in binary format.)
Allow output to be saved as a PLY file and loaded later, rather than just visualized.

Most of these issues are relatively straightforward fixes, if enough time were available; there are no glaring architectural problems with the program that I'm aware of.

Results

	A single range image, read from a PLY file. Data taken from Stanford's repository.
	A single range surface, generated from the previous range image.
	The volume generated from the above range image. Resolution is 25 * 25 * 25 voxels.
	Multiple aligned range surfaces. Each range surface is a different color.
	The result of merging the volumes generated from the preceding six range images.
	The surface generated from the above volume using Marching Cubes.
	The surface generated from the above volume using Marching Cubes on a 50 * 50 * 50 volume. Some parts are more apparent at a higher resolution, especially the back leg; but it's apparent that there are still some problems with volume intersection calculations.

More results, obtained using a cutoff of 0.05 at a resolution of 25 * 25 * 25:

make && ./vrip --res 25 25 25 --cutoff 0.05 data/bunny/data/bun.conf data/bunny/data/bun000.ply data/bunny/data/bun045.ply data/bunny/data/bun090.ply data/bunny/data/bun180.ply data/bunny/data/bun270.ply data/bunny/data/bun315.ply

Even more results, obtained using a cutoff of 0.05 at a resolution of 50 * 50 * 50:

make && ./vrip --res 50 50 50 --cutoff 0.05 data/bunny/data/bun.conf data/bunny/data/bun000.ply data/bunny/data/bun045.ply data/bunny/data/bun090.ply data/bunny/data/bun180.ply data/bunny/data/bun270.ply data/bunny/data/bun315.ply

Resources

The original VRIP paper: A Volumetric Method for Building Complex Models from Range Images; Brian Curless and Marc Levoy
Range image data (bunny): The Stanford 3D Scanning Repository
Greg Turk's PLY Library
Marching Cubes implementation; Lewiner et al.