UrbanScan: 3D modeling of urban scenes

  • "Datasets for PPR"
  • "How StereoScan Works"
  • "PPR Reconstructions"

Point Cloud 3D Reconstruction

The tradicional 3D reconstruction algorithms receive a stereo image sequence acquired by a calibrated stereo pair and give as output the 3D point cloud from the scene. In this sample, we show a mesh constructed using 3D point clouds obtained through a stereo images captured by ZED Stereo Camera from Stereolabs.

    Dataset: Sé Velha - University

Brief intro of Piecewise Planar 3D Reconstruction

Receiving as input an image sequence acquired by a calibrated stereo pair and outputs the camera motion and 3D planes in the scene.

Current multi-view stereo methods still face difficulties in handling situations of:

  • Weak or repetitive texture;
  • Variable illumination;
  • High surface slant;

Piecewise-Planar Reconstruction (PPR) methods are employed to overcome these issues. This new method is based on planes detection, thus, unlike the 3D Reconstruction methods based on point cloud, we describe a 3D model using much less information. As an example, we take a facade of a building as just a plane (3 XYZ-coordinates) with cloud points methods we would have million points to describe it and three XYZ-coordinate values per single point.

Also, estimating the motion from plane correspondences is advantageous since plane features are much less numerous than point features. Man-made environments are often dominated by large size planes, leading to:

  • Fast correspondence and scalability;
  • Correspondence across wide baseline images;
  • Resilience to dynamic foreground;

Reconstruction Results

  • Outdoor Frontal
  • Dataset: Loop in Ladeira do Seminario

    Half the images were used - Loop in Ladeira do Seminario

    All images used - Loop in Ladeira do Seminario

  • Indoor Bumblebee
  • Dataset: ISR Building Entrance

  • Outdoor Bumblebee
  • Dataset: Loop in Condeixa