High Resolution Surface Reconstruction from Overlapping Multiple-Views
Extracting a computer model of a real scene from a sequence of
views, is one of the most challenging and fundamental problems in computer vision.
Stereo vision algorithms allow us to extract from the images a sparse 3D point cloud
on the scene surfaces. However, computing an accurate mesh of the scene based on
such poor quality data points (noise, sparsity) is very difficult. Here we describe
a simple yet original approach that uses both the stereo vision extracted point
cloud and the calibrated images. Our method is a three-stage process in which the
first stage merges, filters and smoothes the input 3D points. The second stage builds
for each calibrated image a triangular depth-map and fuses the set of depth-maps
into a triangle soup that minimize violations of size and visibility constraints.
Finally, a mesh is computed from the triangle soup using a reconstruction method
that combines restricted Delaunay triangulation and Delaunay refinement.
The video on the right presents the described approach.
The short paper abstract below gives references and an overview of the algorithm described in