Texture Compression for Triangle Meshes

By

Hayley Iben
Marci Meingast
Ali Lakhia

Background

• 3D city model generation
• Truck loaded with:
• Laser scanners
• Digital cameras
• Computers to capture data

• Large amount of data per city block
• Mesh: 100,000 – 300,000 triangles
• Texture: 20 – 48 MB
• Interactive walk-thru hard
• Need hardware acceleration
• Throw away four-fifths of texture
• Simplify meshes
• Only one block at a time

Motivation

• Texture data is the bottleneck:
• Graphics card video memory is limited
• Reduce the amount of image data and still be able to reproduce the image accurately
• Enable walk-thru of more data since more texture memory is available

Overview

• Compress the texture by using the repetitious nature of buildings
• Find similar texture through comparing the triangles (which are very small)
• Only keep the blocks of texture used by the triangles, throw out the rest of the data
• Use the new triangle mapping and smaller texture to display entire image

Preprocessing Stage

• Given the original triangle mesh and original image:
• Convert the triangle coordinates to image coordinates
• Pad the image so that data is not lost on the edges by the Gaussian pyramid calculation
• Determine Gaussian pyramid of the image

Algorithm

• For each triangle:
• Look for a match in the smallest Gaussian level by checking every possibility
• If there is a match, check the next pyramid level.
• Continue through the levels until there is a match at every level of the Gaussian pyramid
• Otherwise, continue to the next possible triangle location
• At each level, we are determining the triangle coordinate in that image and checking for a match.

• Triangle comparison:
• Use a polygon fill algorithm to compare each pixel value within the triangle bounds
• Special cases include testing a single point and a line
• Measurement of the match is the squared color distance between all of the pixels

• After all of the triangles in the image have been processed for a match:
• Relevant blocks of texture are moved to form a new texture
• Triangle coordinates are updated to their new location and converted back to the correct format
• Remaining texture data is discarded

Results

• Before



• After



Note: 5,000 of 29,000 triangles moved

• Before



• After

• Before



• After



Note: Texture decreased by 8.7%

Future Directions

• Try different traversal of Gaussian pyramid
• Give preference to certain blocks to avoid scattering
• Pick starting blocks using RANSAC
• Optimization of parameters