| Detection Project Overview | Anticipation Project Overview | Data/Code | Results |

Cornell Activity Datasets: CAD-60 & CAD-120

  • Data
    • Raw Data
  • Data Format
    • Skeleton Data Format
    • RGBD Data Format
  • Code
    • Feature Extraction
    • Skeleton Visualization Tool
activity sample images

Data

The CAD-60 and CAD-120 data sets comprise of RGB-D video sequences of humans performing activities which are recording using the Microsoft Kinect sensor.

CAD-60

CAD-60 dataset features:

  • 60 RGB-D videos
  • 4 subjects: two male, two female, one left-handed
  • 5 different environments: office, kitchen, bedroom, bathroom, and living room
  • 12 activities: rinsing mouth, brushing teeth, wearing contact lens, talking on the phone, drinking water, opening pill container, cooking (chopping), cooking (stirring), talking on couch, relaxing on couch, writing on whiteboard, working on computer
  • tracked skeletons
Each video come with RGB images, Depth images, and the tracked skeletons.

Information: README Sample Images State of the art results
Download
RGB-D + Skeleton:
Person 1 Person 2 Person 3 Person 4

CAD-120

CAD-120 dataset features:

  • 120 RGB-D videos of long daily activities
  • 4 subjects: two male, two female, one left-handed
  • 10 high-level activities: making cereal, taking medicine, stacking objects, unstacking objects, microwaving food, picking objects, cleaning objects, taking food, arranging objects, having a meal
  • 10 sub-activity labels: reaching, moving, pouring, eating, drinking, opening, placing, closing, scrubbing, null
  • 12 object affordance labels: reachable, movable, pourable, pourto, containable, drinkable, openable, placeable, closable, scrubbable, scrubber, stationary
  • tracked skeletons
Click here for samle images.


RGB-D images: Person 1 Person 2 Person 3 Person 4 README
RGB-D text data: Person 1 Person 2 Person 3 Person 4 README
Annotations: Person 1 Person 2 Person 3 Person 4 README
Features: Features README

Data Format

Skeleton Data Format

Skeleton data consists of 15 joints. There are 11 joints that have both joint orientation and joint position. And, 4 joints that only have joint position. Each row follows the following format.

  Frame#,ORI(1),P(1),ORI(2),P(2),...,P(11),J(11),P(12),...,P(15)
    
    Frame# => integer starting from 1
    
    ORI(i) => orientation of ith joint
                0 1 2
                3 4 5
                6 7 8
              3x3 matrix is stored as followed by CONF
                0,1,2,3,4,5,6,7,8,CONF
              Read NITE PDF (see below) to get more detail about the matrix
                
    P(i)   => position of ith joint followed by CONF
                x,y,z,CONF
              values are in milimeters
              
    CONF   => boolean confidence value (0 or 1)
              Read NITE PDF (see below) to get more detail about the confidence value
    
  Joint number -> Joint name
     1 -> HEAD
     2 -> NECK
     3 -> TORSO
     4 -> LEFT_SHOULDER
     5 -> LEFT_ELBOW
     6 -> RIGHT_SHOULDER
     7 -> RIGHT_ELBOW
     8 -> LEFT_HIP
     9 -> LEFT_KNEE
    10 -> RIGHT_HIP
    11 -> RIGHT_KNEE
    12 -> LEFT_HAND
    13 -> RIGHT_HAND
    14 -> LEFT_FOOT
    15 -> RIGHT_FOOT
    

Read page 10~13 of NITE 1.3 PDF for more detail on skeleton orientation, position, and confidence values.

RGBD Data Format

RGBD data has resolution of 240 by 320. RGB is saved as three-channel 8-bit PNG file. And, Depth is saved as single-channel 16-bit PNG file. Due to alignment of Depth and RGB data, some pixels on the edges will have value of 0. Refer to "Feature Extraction" in Sung et al. code below to look at how to parse these PNG files.

RGBD sample image..

Code

All of the code described in our papers can be retrieved by running following command.

  1. Sung et al. AAAI PAIR 2011, ICRA 2012

     
            git clone git://github.com/jysung100/activity_detection.git 
         or   
            download(zip)
            

  2. Activity Labeling Code, Koppula et al. IJRR 2013

            git clone git://github.com/hemakoppula/human_activity_labeling.git
            

    The human_activity_labeling repository has code for the following:

    • Feature generation code: ROS and PCL packages.
    • Learning and inference code (using python interface to svm_struct).

  3. Activity Anticipation Code, Koppula and Saxena. RSS 2013

            git clone git://github.com/hemakoppula/human_activity_anticipation.git
            

    The human_activity_anticipation repository has code for the following:

    • Anticipation code: PCL package.
    • Learning code (using python interface to svm_struct).

Skeleton Visualization Tool

Available in Sung et al. code repository above.

Turns specific frame of skeleton data into Matlab 3D plot. It shows each joints location as well as orientation of joints that have such information. Refer to README for more details.
Requires: Matlab (tested on R2010a)

Matlab skeleton visualization sample image..