A reference implementation for the paper Texture Defragmentation for Photo-Reconstructed 3D Models by Andrea Maggiordomo, Paolo Cignoni and Marco Tarini.
We propose a method to improve an existing parametrization (UV-map layout) of a textured 3D model, targeted explicitly at alleviating typical defects afflicting models generated with automatic photo-reconstruction tools from real-world objects. This class of 3D data is becoming increasingly important thanks to the growing popularity of reliable, ready-to-use photogrammetry software packages. The resulting textured models are richly detailed, but their underlying parametrization typically falls short of many practical requirements, particularly exhibiting excessive fragmentation and consequent problems. Producing a completely new UV-map, with standard parametrization techniques, and then resampling a new texture image, is often neither practical nor desirable for at least two reasons: first, these models have characteristics (such as inconsistencies, high resolution) that make them unfit for automatic or manual parametrization; second, the required resampling leads to unnecessary signal degradation because this process is unaware of the original texel densities. In contrast, our method improves the existing UV-map instead of replacing it, balancing the reduction of the map fragmentation with signal degradation due to resampling, while also avoiding oversampling of the original signal. The proposed approach is fully automatic and extensively tested on a large benchmark of photo-reconstructed models; quantitative evaluation evidences a drastic and consistent improvement of the mappings.
Dependencies:
- QT5
Building using qmake on Linux with GCC should be straightforward:
mkdir build && cd build
qmake ../texture-defrag/texture-defrag.pro -spec linux-g++
make
The executable takes as arguments the input mesh file and optional parameters to control the various steps of the texture defragmentation algorithm.
Usage: ./texture-defrag MESHFILE [-mbdgutao]
MESHFILE specifies the input mesh file (supported formats are obj, ply and fbx)
-m <val> Matching error tolerance when attempting merge operations. (default: 2)
-b <val> Maximum tolerance on the seam-length to chart-perimeter ratio when attempting merge operations. Range is [0,1]. (default: 0.2)
-d <val> Local ARAP distortion tolerance when performing the local UV optimization. (default: 0.5)
-g <val> Global ARAP distortion tolerance when performing the local UV optimization. (default: 0.025)
-u <val> UV border reduction target in percentage relative to the input. Range is [0,1]. (default: 0)
-a <val> Alpha parameter to control the UV optimization area size. (default: 5)
-t <val> Time-limit for the atlas clustering (in seconds). (default: 0)
-o <val> Output mesh file. Supported formats are obj and ply. (default: out_MESHFILE)
-l <val> Logging level. 0 for minimal verbosity, 1 for verbose output, 2 for debug output. (default: 0)
@article {maggiordomo2021defragmentation,
journal = {Computer Graphics Forum},
title = {{Texture Defragmentation for Photo-Reconstructed 3D Models}},
author = {Maggiordomo, Andrea and Cignoni, Paolo and Tarini, Marco},
year = {2021},
publisher = {The Eurographics Association and John Wiley & Sons Ltd.},
ISSN = {1467-8659},
DOI = {10.1111/cgf.142615}
}