MPII Multi-Kinect Dataset

---------------------------------------------------------
License
---------------------------------------------------------

The data is only to be used for scientific purposes and must not be republished other than by Max Planck Institue for Informatics. The scientific use includes processing and showing it in publications and presentations. When using it please cite [1].

---------------------------------------------------------
Files
---------------------------------------------------------
This dataset consists of:
pair_image: 2D images captured with four Kinect cameras using OpenKinect software (PPM (color) and PGM (depth)),
ground_truth: ground truth data,
pcd_data: 3D point cloud data for four Kinect camera,
pcd_datasm_reg: 3D point cloud data registered (using depth-averaged point cloud),
script_and_calibration: includes evaluation script and calibration data,
README.txt

---------------------------------------------------------
Dataset size:
---------------------------------------------------------
pcd_data:10 G
image: 4.5 G
pcd_reg:0.5 G
---------------------------------------------------------

---------------------------------------------------------
Ground truth
---------------------------------------------------------
The data are recorded with four different Kinect cameras simultaneously.
In our experiments, we use one camera (top camera / kin2) as ground truth and the three remaining for detection.

The ground truth data is available in two types: bounding box (annotation_bbox) and polygons (annotation_polygon).

---------------------------------------------------------
Calibration
---------------------------------------------------------
The calibration matrix in calibration text files consist of extrinsic matrices between two camera views.
These are estimated with ICP in point cloud library.
calibration file explanation:
for file "fincalmat12.data", the extrinsic matrix transforms point cloud of kin1 view to point cloud kin2 view, and so on.

There two calibration files:
1. calibration A:
-datasetkit
-scenekit 1, 2, 3
-background_A

2. calibration B:
-scenekit 4, 5
-background_B

For 3D reconstruction from the depth image we follow OpenKinect formulas:
w = 640
h = 480
minDistance = -10
scaleFactor = 0.0021

X =  (x - (w/2)) * (z + minDistance) * scaleFactor * (w/h);
Y =  (y - (h/2)) * (z + minDistance) * scaleFactor;
Z = z;

---------------------------------------------------------
Evaluation scripts
---------------------------------------------------------
centerpoint_eval = function to compute distance euclidean distance between prediction and ground truth (in script_and_calibration)
backproject_2dpoints = function to compute 3D points from 2D points
read_depth_img = function to read depth image

---------------------------------------------------------
References
---------------------------------------------------------
[1] 3D Object Detection with Multiple Kinects,
W. Susanto, M. Rohrbach and B. Schiele,
2nd Workshop on Consumer Depth Cameras for Computer Vision in conjunction with ECCV 2012, Firenze, Italy, (2012)