RGB-D People Dataset
Annotated people and tracks in RGB-D Kinect data
Description
This dataset contains 3000+ RGB-D frames acquired in a university hall from three vertically mounted Kinect sensors.
The data contains mostly upright walking and standing persons seen from different orientations and with different levels of occlusions.
This dataset has been re-annotated with respect to [1][2] and it contains minor annotation differences.
CLARIFICATION: Results in paper [1] are computed each three frames. Results in paper [2] are computed in each frame.
CLARIFICATION 2: Some people experienced problems in loading the depth images. In case that depth values are over 10^4, then swap the position of the two bytes for each depth pixel. It is a big/little endian interpretation problem.
If you are considering using this data, please reference the following:
The dataset can be download from http://www.informatik.uni-freiburg.de/~spinello/sw/rgbd_people_unihall.tar.gz
Three vertically mounted kinect have been assembled on a sensory tower at ~1.50m height. The setup has been chosen to minimize the IR projector cross-talk between the sensors. Data has been acquired @ 30Hz.
Each annotated track is defined in a file named as TRACK_000N.DAT . The file describes the position of the target in each depth frame it appears. The target visibility status is also annotated. The track-ID is encoded in the file name (TRACK_000N.DAT -> track-ID N). The content of a track file is as following:
where:
Visual explaination:
Depth images are saved as 16bits, 1 channel PGM images [640x480]. They contain the raw data content from the Kinect sensor. Namely, each pixel has value between [0,1084]. Please check this page in case you are interested in metric values.
Some people experienced problems in loading the depth images. In case that depth values are over 10^4, then swap the position of the two bytes for each depth pixel. It is a big/little endian interpretation problem.
RGB images are saved as 8bits, 3 channels PPM images [640x480].
RGB intrinsics params:
Depth intrinsics params:
Extrinsic params:
Copyright (c) 2011 Luciano Spinello, Matthias Luber and Kai O. Arras
This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported License.
This dataset has been re-annotated with respect to [1][2] and it contains minor annotation differences.
CLARIFICATION: Results in paper [1] are computed each three frames. Results in paper [2] are computed in each frame.
CLARIFICATION 2: Some people experienced problems in loading the depth images. In case that depth values are over 10^4, then swap the position of the two bytes for each depth pixel. It is a big/little endian interpretation problem.
Citations
If you are considering using this data, please reference the following:
- [1] L. Spinello, K. O. Arras. People Detection in RGB-D Data. IEEE Int. Conf. on Intelligent Robots and Systems (IROS), 2011.
- [2] M. Luber and L. Spinello and K. O. Arras People Tracking in RGB-D Data With On-line Boosted Target Models. IEEE Int. Conf. on Intelligent Robots and Systems (IROS), 2011.
Download
The dataset can be download from http://www.informatik.uni-freiburg.de/~spinello/sw/rgbd_people_unihall.tar.gz
Sensory Setup
Three vertically mounted kinect have been assembled on a sensory tower at ~1.50m height. The setup has been chosen to minimize the IR projector cross-talk between the sensors. Data has been acquired @ 30Hz.
Annotation Format
Each annotated track is defined in a file named as TRACK_000N.DAT . The file describes the position of the target in each depth frame it appears. The target visibility status is also annotated. The track-ID is encoded in the file name (TRACK_000N.DAT -> track-ID N). The content of a track file is as following:
IMAGE_NAME_ID TS X_tl_dpt Y_tl_dpt WDT_dpt HGT_dpt X_tl_rgb Y_tl_rgb WDT_rgb HGT_rgb VSB
...
IMAGE_NAME_ID = image filename, seq0_0001_2.pgm [ID: camera id, 0 = left, 1 = center, 2 = right] TS = timestamp X_tl_dpt, Y_tl_dpt = topleft coord. bounding box in the depth image WDT_dpt, HGT_dpt = size of the bounding box in the depth image X_tl_rgb, Y_tl_rgb = topleft coord. bounding box in the RGB image WDT_rgb, HGT_rgb = size of the bounding box in the RGB image VSB = visibility in the depth image [0 = hidden 1 = fully visible, 2 = partially visible ] TS = timestamp
Depth images are saved as 16bits, 1 channel PGM images [640x480]. They contain the raw data content from the Kinect sensor. Namely, each pixel has value between [0,1084]. Please check this page in case you are interested in metric values.
Some people experienced problems in loading the depth images. In case that depth values are over 10^4, then swap the position of the two bytes for each depth pixel. It is a big/little endian interpretation problem.
RGB images are saved as 8bits, 3 channels PPM images [640x480].
Calibration
RGB intrinsics params:
k1=2.6451622333009589e-01 k2=-8.3990749424620825e-01 k3=9.1192465078713847e-01 p1=-1.9922302173693159e-03 p2=1.4371995932897616e-03 fx=5.2921508098293293e+02 fy=5.2556393630057437e+02 cx=3.2894272028759258e+02 cy=2.6748068171871557e+02
k1=-2.6386489753128833e-01 k2=9.9966832163729757e-01 k3=-1.3053628089976321e+00 p1=-7.6275862143610667e-04 p2=5.0350940090814270e-03 fx=5.9421434211923247e+02 fy=5.9104053696870778e+02 cx=3.3930780975300314e+02 cy=2.4273913761751615e+02
T= [1.9985242312092553e-02 -7.4423738761617583e-04 -1.0916736334336222e-02]
{rotation angles} roll=1.22775236635173e-02 pitch=-1.74713102421449e-02 yaw=-1.47813574129019e-03
Copyright and License Information
Copyright (c) 2011 Luciano Spinello, Matthias Luber and Kai O. Arras
This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported License.
This dataset is distributed WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.