MAP Visibility Estimation for
Large-Scale Dynamic 3D Reconstruction
(Dataset)

Hanbyul Joo, Hyun Soo Park, and Yaser Sheikh

Carnegie Mellon University

Paper to Reference

Hanbyul Joo, Hyun Soo Park, and Yaser Sheikh. MAP Visibility Estimation for Large-Scale Dynamic 3D Reconstruction, In CVPR, 2014. [Paper, BibTeX, Slide (pdf), Poster (pdf), Project page]

Description

This dataset is captured by the Panoptic Studio at Carnegie Mellon University using 480 cameras.
The Panoptic Studio has 20 panels, and each panel has 24 VGA resolution cameras.
Each panel has a panel index from 1 to 20, and a camera in each panel has a local camera index from 1 to 24. Thus, a camera can be identified by a pair of panel index and camera index, as (panelIdx)_(cameraIdx). For example, 01_24 is a camera in panel 1 with camera index 24.
Note that both panel and camera indices are not ordered. That is, panel 1 and panel 2 may not be neighboring panels. However, you may assume that cameras belongs to a same panel are located nearby. The "geodesic" panel order is as follows.
-1-3-17-5-4-6-18-8-7-9-19-11-10-12-20-14-13-15-16-2-1-
We use the following naming rule for the capture images.

(8 digit of frame number)_(panelIdx)_(camIdx).png
For example, 00000314_04_07.png is the 314th frame of the 7th camera in the 4th panel.

Calibration

Download

Calibration parameter files have same the naming rule (panelIdx_camIdx) described above.
Calibration parameter files for each camera are composed of two files, intrinsic and extrinsic.
For example "04_13.txt" is the intrinsic calibration file for the 13th camera in the 4th panel, and "04_13_ext.txt" is the extrinsic calibration file for the same camera.
The calibration parameters are generated by the VisualSFM by Changchang Wu.

Intrinsic Parameter: (panelIdx)_(camIdx).txt

Composed of 11 floating point numbers as follows.
K11 K12 K13 K21 K22 K23 K31 K32 K33 distortionParam1 distortionParam2 xxx
All the sequences contain rectified images, thus both distortion parameters are zeros.

Extrinsic Parameter: (panelIdx)_(camIdx)_ext.txt

Composed of 7 floating point numbers
The first 4 numbers are quaternion representation of R
The last 3 numbers are the camera center, which is -invR x t
Thus, t = -R*Center

Dynamic 3D Scene reconstruction

The dynamamic 3D reconstruction results, what we call the "Trajectory stream", are generated by the method presented in our CVPR 2014 paper.
Note this description for the the exported trajectory stream file format.
You can load and visualize trajectory strems using this viewer in MATLAB.