Jack Rabbot Dataset and Benchmark (JRDB)

JRDB is the largest benchmark dataset for 2D-3D person tracking, including:
  • Over 60K frames (64 minutes) sensor data captured from 5 stereo cylindrical 360° RGB cameras and two LiDAR sensors
  • 54 sequences from different indoor and outdoor locations in the Stanford university campus
  • Over 2.3 milion high quality 2D bounding box annotations on 360° cylindrical video streams generated from 5 stereo cameras
  • Over 1.8 milion high quality 3D oriented bounding box annotations on continuous 3D point clouds from two Velodyne 16 LiDARs
  • More than 3500 time consistent trajectories (tracks) in 2D and 3D
  • Download train and test splits.


    The most relevant parameters and information about the sensor setup is included in the following document:
  • Sensor Setup
  • The dataset file structure and the format of the labels (fields, order, ...) is described in the document below. The Preparing Submissions site includes development kits to work with the dataset, develop and test your solution.
  • File Structure and Label Format
  • JRDB White Paper

    We have developed and published a white paper including the most relevant information about the JRDB: the sensor suite used to record the data, the most relevant specs and an analysis of the type of data and annotations included.