This repository shows how to work with the Panoptic Studio (Panoptic) data.

Follow these steps to set up a simple example:

git clone https://github.com/CMU-Perceptual-Computing-Lab/panoptic-toolbox
cd panoptic-toolbox

To download a dataset, named "171204_pose1_sample" in this example, run the following script.

./scripts/getData.sh 171204_pose1_sample

This bash script requires curl or wget.

This script will create a folder "./171204_pose1_sample" and download the following files.

• 171204_pose1_sample/hdVideos/hd_00_XX.mp4 #synchronized HD video files (31 views)
• 171204_pose1_sample/vgaVideos/KINECTNODE%d/vga_XX_XX.mp4 #synchrponized VGA video files (480 views)
• 171204_pose1_sample/calibration_171204_pose1_sample.json #calibration files
• 171204_pose1_sample/hdPose3d_stage1_coco19.tar #3D Body Keypoint Data (coco19 keypoint definition)
• 171204_pose1_sample/hdFace3d.tar #3D Face Keypoint Data
• 171204_pose1_sample/hdHand3d.tar #3D Hand Keypoint Data

Note that this sample example currently does not have VGA videos.

You can also download any other seqeunce through this script. Just use the the name of the target sequence instead of the "171204_pose1_sample".

For example,

./scripts/getData.sh 171204_pose1

for the full version of 171204_pose1 sequence.

You can also specify the number of videos you want to donwload.

./scripts/getData.sh (sequenceName) (VGA_Video_Number) (HD_Video_Number)

For example, the following command will download 240 vga videos and 10 videos.

./scripts/getData.sh 171204_pose1_sample 240 10

Note that we have sorted the VGA camera order so that you download uniformly distributed view.

You can find the list of currently available sequences in the following link:

List of release sequences

You can see the example videos and other information of each sequence in our website:

Browsing dataset.

Check the 3D viewer in each sequence page where you can visualize 3D skeletons in your web browser. For example:

http://domedb.perception.cs.cmu.edu/171204_pose1.html

This step requires ffmpeg.

./scripts/extractAll.sh 171204_pose1_sample

This will extract images, for example 171204_pose1_sample/hdImgs/00_00/00_00_00000000.jpg, and the corresponding 3D skeleton data, for example 171204_pose1_sample/hdPose3d_stage1_coco19/body3DScene_00000000.json.

extractAll.sh is a simple script that combines the following set of commands (you shouldn't need to run these again):

cd 171204_pose1_sample
../scripts/vgaImgsExtractor.sh # PNG files from VGA video (25 fps)
../scripts/hdImgsExtractor.sh # PNG files from HD video (29.97 fps)
tar -xf vgaPose3d_stage1.tar # Extract skeletons at VGA framerate
tar -xf hdPose3d_stage1.tar # Extract skeletons for HD
cd ..

This codes require numpy, matplotlib.

Visualizing 3D keypoints (body, face, hand):

cd python
jupyter notebook demo_3Dkeypoints_3dview.ipynb

The result should look like this.

Reprojecting 3D keypoints (body, face, hand) on a selected HD view:

cd python
jupyter notebook demo_3Dkeypoints_reprojection_hd.ipynb

The result should look like this.

Note: Matlab code is outdated, and does not handle 3D keypoint outputs (coco19 body, face, hand). Please see this code only for reference. We will update this later.

Matlab example (outdated):

>>> cd matlab
>>> demo

We reconstruct 3D skeleton of people using the method of Joo et al. 2018.

The output of each frame is written in a json file. For example,

{ "version": 0.7, 
"univTime" :53541.542,
"fpsType" :"hd_29_97",
"bodies" :
[
{ "id": 0,
"joints19": [-19.4528, -146.612, 1.46159, 0.724274, -40.4564, -163.091, -0.521563, 0.575897, -14.9749, -91.0176, 4.24329, 0.361725, -19.2473, -146.679, -16.1136, 0.643555, -14.7958, -118.804, -20.6738, 0.619599, -22.611, -93.8793, -17.7834, 0.557953, -12.3267, -91.5465, -6.55368, 0.353241, -12.6556, -47.0963, -4.83599, 0.455566, -10.8069, -8.31645, -4.20936, 0.501312, -20.2358, -147.348, 19.1843, 0.628022, -13.1145, -120.269, 28.0371, 0.63559, -20.1037, -94.3607, 30.0809, 0.625916, -17.623, -90.4888, 15.0403, 0.327759, -17.3973, -46.9311, 15.9659, 0.419586, -13.1719, -7.60601, 13.4749, 0.519653, -38.7164, -166.851, -3.25917, 0.46228, -28.7043, -167.333, -7.15903, 0.523224, -39.0433, -166.677, 2.55916, 0.395965, -30.0718, -167.264, 8.18371, 0.510041]
}
] }

Here, each subject has the following values.

id: a unique subject index within a sequence. Skeletons with the same id across time represent temporally associated moving skeletons (an individual). However, the same person may have multiple ids if he/she appeared multiple different times (e.g., left the studio and reappeared later).

joints19: 19 3D joint locations, formatted as [x1,y1,z1,c1,x2,y2,z2,c2,...] where each c is a per-joint confidence score.

The 3D skeletons have the following keypoint order:

0: Neck
1: Nose
2: BodyCenter (center of hips)
3: lShoulder
4: lElbow
5: lWrist,
6: lHip
7: lKnee
8: lAnkle
9: rShoulder
10: rElbow
11: rWrist
12: rHip
13: rKnee
14: rAnkle
15: rEye
16: lEye
17: rEar
18: lEar

Note that this is different from OpenPose output order, although our method is based on it.

Note that we used to use an old format (named mpi15 as described in our outdated document), but we do not this format anymore.

Kinoptic Studio is a subsystem of Panoptic Studio, which is composed of 10 Kinect2 sensors.

Kinoptic Studio can be independently used from the Panoptic Studio.

See our PtCloudDB document for more details

Follow these steps to set up a simple example:

Assuming you want to donwload a sequence named "160422_haggling1"

./scripts/getData_kinoptic.sh 160422_haggling1

This script will download the following files.

• 160422_haggling1/kinect_shared_depth/ksynctables.json #sync table
• 160422_haggling1/kinect_shared_depth/KINECTNODE%d/depthdata.dat #depth files
• 160422_haggling1/kcalibration_160422_haggling1.json #multiple kinects calibration files
• 160422_haggling1/kinectVideos/kinect_50_%d.mp4 #rgb video files

cd 160422_haggling1
../scripts/kinectImgsExtractor.sh

matlab ./matlab/demo_kinoptic_gen_ptcloud.m

Note that you should set your "root_path" and "seqName" in this demo file.

matlab ./matlab/demo_kinoptic_projection.m

Similarly, note that you should set your "root_path" and "seqName" in this demo file.

You can download all sequences included in our 3D PointCloud DB ver.1 using the following script:

./scripts/getDB_ptCloud_ver1.sh

Panoptic Studio Dataset is freely available for free non-commercial use.

By using the dataset, you agree to cite at least one of the following papers.

@InProceedings{Joo_2015_ICCV, author = {Joo, Hanbyul and Liu, Hao and Tan, Lei and Gui, Lin and Nabbe, Bart and Matthews, Iain and Kanade, Takeo and Nobuhara, Shohei and Sheikh, Yaser}, title = {Panoptic Studio: A Massively Multiview System for Social Motion Capture}, booktitle = {ICCV}, year = {2015} }

@article{Joo_2017_TPAMI, title={Panoptic Studio: A Massively Multiview System for Social Interaction Capture}, author={Joo, Hanbyul and Simon, Tomas and Li, Xulong and Liu, Hao and Tan, Lei and Gui, Lin and Banerjee, Sean and Godisart, Timothy Scott and Nabbe, Bart and Matthews, Iain and Kanade, Takeo and Nobuhara, Shohei and Sheikh, Yaser}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, year={2017} }

@article{Simon_2017_CVPR, title={Hand Keypoint Detection in Single Images using Multiview Bootstrapping}, author={Simon, Tomas and Joo, Hanbyul and Sheikh, Yaser}, journal={CVPR}, year={2017} }