UCL Logo VISUAL COMPUTING

README

===========================================================================
====  Scalable Inside-Out Image-Based Rendering - Dataset Release      ====
====  Peter Hedman, Tobias Ritschel, George Drettakis, Gabriel Brostow ====
====  SIGGRAPH Asia 2016                                               ====
===========================================================================

  Structure-from-Motion is ambiguous with respect to orientation and scale. This means 
  that our SfM reconstructions do not have consitent scales and up-vectors. We automatically
  compute scale from the (metric) RGB-D images from the depth scanner. For easier navigation,
  we manually estimated the up vector forward vector in the scene.

  While we store our depth maps as 16-bit PNGs with a millimeter scale, everything else
  (all the meshes and the camera poses) live in the (arbitrary) coordinate space produced 
  by the SfM tool.

  Use the following code to convert load a depth map with the scale of the SfM reconstruction:

      // mm_to_sfm can be found in parameters.txt
      cv::Mat load_depth_map(const std::string& path, float mm_to_sfm) {
          static const float MAX_16BIT_VALUE = 65535.f;
          cv::Mat input = cv::imread(path);
          cv::Mat output;
          input.convertTo(output, CV_32F, mm_to_sfm / MAX_16BIT_VALUE);
          return output;
      }

==========================================
== Scene parameters (in parameters.txt) ==
==========================================

    estimated_up      : Aproximate up vector expressed in SfM coordinates.
    estimated_forward : Approximate forward vector expressed in SfM coordinates.
    mm_to_sfm         : Scaling factor which converts from a millimeter scale to the scale

  Note that the up and forward vectors are not necessary. However, they are convenient 
  if you want the horizon to be correct when you're navigating the scene.

================
== Input data ==
================

    Input_Data\Camera_Images_High_Res\*.jpg    : Our input images after pre-processing
                                                 (cropped, denoised, and white-balanced).
    Input_Data\RGBD_Scan.oni                   : The complete RGB-D scan, stored in the 
                                                 OpenNI file format.
    Input_Data\RGBD_Scan_As_Images\color-*.jpg : Color images extracted the RGB-D scan.
    Input_Data\RGBD_Scan_As_Images\depth-*.png : Depth images extracted the RGB-D scan,
                                                 stored in millimeters as 16-bit PNGs.

  Note that the color and depth videos from the RGB-D sensor are *NOT SYNCED*. To overcome
  this, we only extracted color images that have a corresponding depth frame (within 60 ms).

========================
== SfM reconstruction ==
========================

  We store camera poses for both the RGB-D images and the digital camera images in the NVM
  file format, which can be read by VisualSFM (http://ccwu.me/vsfm/).

    SfM\Input\Camera_Images_Low_Res\*.jpg                : Input images downscaled by 4x,
                                                           to conserve memory and speed up
                                                           reconstruction.
    SfM\Input\RGBD_Scan_As_Images_Subsampled\color-*.jpg
    SfM\Input\RGBD_Scan_As_Images_Subsampled\depth-*.png : We subsample the RGB-D scan to 
                                                           fewer than 500 images. This also
                                                           speeds up reconstruction.

    SfM\Output\camera_poses.nvm   : The camera poses for both the RGB-D images and 
                                    the digital camera images. Stored as an NVM file.
    SfM\Output\images\color-*.jpg : The color components of the RGB-D images,
                                    after after correcting for radial distortion.
    SfM\Output\images\*.jpg       : The digital camera images, after correcting for
                                    radial distortion.

===========================
== Global reconstruction ==
===========================

    Global_Reconstruction\Fused_Point_Cloud_From_RGBD_Images.ply : All the RGB-D images fused
                                                                   into a point cloud with
                                                                   colors and normals. Stored
                                                                   as PLY files (binary).
    Global_Reconstruction\Fused_Mesh_From_RGBD_Images.ply        : A dense mesh reconstructed 
                                                                   from the point cloud. Stored
                                                                   as PLY files (binary).

  Note that both of these reconstructions use the same coordinate space as the 
  SfM reconstruction above. We use Screened Poisson Surface Reconstruction
      (http://www.cs.jhu.edu/~misha/Code/PoissonRecon/Version5/)
  to convert the point cloud into a dense mesh.

  Both reconstructions can be loaded with e.g. MeshLab (http://www.meshlab.net/).
      
=============================
== Per-view reconstruction ==
=============================

    Per_View_Reconstruction\Depth_Maps_For_Camera_Images\*.png : Our refined depth maps for 
                                                                 the digital camera images.
                                                                 Stored in millimeters as
                                                                 16-bit PNGs.
    Per_View_Reconstruction\Meshes_For_Camera_Images\*.ply     : Our local meshes for 
                                                                 the digital camera images.
                                                                 Stored as PLY files (ASCII)
                                                                 with normals and texture
                                                                 coordinates.

  Note that the meshes live in the SfM coordinate space also used by the global reconstruction.
  The PLY meshes can be loaded with e.g. MeshLab (http://www.meshlab.net/).