Nebula : A Mesh-Error Measuring Tool Based on Rendered Images

Fu-Chung Huang           and           Bing-Yu Chen

National Taiwan University

Abstract:

In this paper we propose a tool that compares the difference of the surface and its simplified representation. Geometry-based metrics usually cannot truly reflect how the user perceives the difference when surfaces are rendered on the screen, hence we provide an alternative tool that is image-based, which complement the deficiency of the geometry-based. Two metrics are implemented in the tool: pixel-wise Mean Squared Image Differences and Cortex Pyramid differences.

We not only implement these image-based metrics but also provide an assessment of the results between image- and geometry- based metrics.


Top

Introduction:

We provide a tool that measures the differences between a triangle mesh and its simplified approximation based on rendered images. Within the tool, two common metrics are implemented, namely Mean Squared Image Difference and Cortex Pyramid.

The basic principle behind this tool is that current public domain tool measuring errors are geometry-based, and we think that the quality of simplified mesh should be evaluated more closely to the human vision model, by using what we perceived on the screen. This work is mainly a complement of Metro[CRS98], and implementation based on what was proposed in [Lin00].

More specific details can be found in our paper here.

Top

Usage:

dir:\> nebula source_file_name target_file_name [-opt]

[-opt] can be the followings:

Top

Implementation Nodes:

The implementation details we adopted from Peter Lindstrom's Ph.D thesis are slightly different, mainly on the setting of parameters. E.g. the Display Transfer Function and the size of Gaussian mask.


Future updates should include:

•  Scalable image size. Currently we only support captured resolution of 512 x 512.
•  Wider system support. Windows system is our current concern, yet for greater research purpose, a portable implementation will be provided in the future.

Top

Downloads:

Windows executable only
Source code (VC6.0 project)
Paper
Testing examples

For executable to run correctly, you will need GLUT library, and for source compliation, you will need to install the Intel OpenCV library.

In order to compile the project, you also need to configure the OpenCV library for your own machine. For specific detail please refer to the OpenCV Documentation.

Top

Contact Information:

jonash_at_cmlab.csie.ntu.edu.tw

Top

Examples:

Bunny
35103v/70202f -> 252v/500f
The MSID measurement : 13.1203
The cortex measurement : 7.75903

Horse
48485v/96966f -> 252v/500f
The MSID measurement : 9.40479
The cortex measurement : 5.56226

Skeleton Hand
327323v/654666f -> 249v/500f
The MSID measurement : 11.9534
The cortex measurement : 4.90993

Scanned Hand
38219v/76438f -> 251v/500f
The MSID measurement : 6.85732
The cortex measurement : 4.07545

Sphere
9902v/19800f -> 52v/100f
The MSID measurement : 9.10971
The cortex measurement : 5.3312

Top

References:

[CRS98] CIGNONI P., ROCCHINI C., SCOPIGNO R.: Metro: Measuring error on simplified surfaces. Comput. Graph. Forum 17, 2 (1998), 167–174.

[Lin00] LINDSTROM P.: Model simplification using image and geometry-based metrics. PhD thesis, 2000. Adviser- Greg Turk.

Top