KTH Moving Objects Dataset

image0

KTH Scitos G5 robot - Rosie

This dataset extends KTH Longterm Dataset Labels with more locations within the same office environment at KTH. The dataset contains a subset of the labels and these objects are consistently located in different positions in multiple rooms. The label subset annotated in this dataset is {chair6, chair1, water_boiler, backpack1, pillow, trash_bin, backpack3, chair2, hanger_jacket, backpack2}. Each observation consists of a set of 17 RGB-D images obtained by moving the pan-tilt in a horizontal pattern, in increments of 20 degrees. In addition to the raw sensor data, each observation contains object annotations (masks and labels). The data is a part of the Strands EU FP7 project.

Dataset structure

The data is structured in folders as follows: YYYYMMDD/patrol_run_YYY/room_ZZZ, where:

  • YYYYMMDD represents the year, month & day when those particular observations were acquired. Each such folder contains the patrol runs the robot collected on that specific date.
  • patrol_run_YYY represents one of the patrol runs collected by the robot.
  • room_ZZZ represents a particular observation collected during a patrol run.

Each folder of the type YYYMMDD/patrol_run_YYY/room_ZZZ contains the following files:

  • room.xml - contains information relevant for the observation (described in the next section)
  • surfel_map.pcd - the surfel map constructed by using the data fusion component of Elastic Fusion
  • complete_cloud.pcd - the point cloud of the observation (obtained by thresholding the confidence of surfel_map.pcd at 0.3)
  • intermediate_cloudXXXX.pcd - ordered point clouds, each corresponding to an RGB and depth image acquired by the camera while conducting the sweep (17 such point clouds for each observation, with XXXX going from 0000 to 0016)
  • rgb_XXXX.jpg - RGB image generated from the corresponding intermediate_cloudXXXX.pcd
  • depth_XXXX.png - depth image generated from the corresponding intermediate_cloudXXXX.pcd
  • rgb_XXXX_label_#.jpg - binary mask corresponding to one of the objects annotated in the image rgb_XXXX.jpg
  • rgb_XXXX_label_#.pcd - point cloud generated from object mask rgb_XXXX_label_#.jpg
  • rgb_XXXX_label_#.txt - object label corresponding to annotation rgb_XXXX_label_#.jpg
  • rgb_XXXX_label_#.xml - xml file storing the labelled object data (created with data from rgb_XXXX_label_#.pcd, rgb_XXXX_label_#.txt)
  • rgb_XXXX_object_#.jpg - rgb mask corresponding to one of the objects annotated in the image rgb_XXXX.jpg
The description of the room.xml file accompanying an observation can be found here.
Each object xml file (rgb_XXXX_label_#.xml) contains the following data:
  • label - object label
  • filename - object pcd file
  • Centroid - object centroid, in the map frame
  • LogTime - time when the object was observed

Parsing

A parser is provided here (can be installed with `` sudo apt-get install ros-indigo-metaroom-xml-parser``) which reads in the data and returns C++ data structures encapsulating the low-level data from the disk. Form more information please refer to the parser README ( or here for a list of supported methods). Information about how to use the Strands package repository can be found here.
Assuming the parser has been installed, the labelled data can be visualized with the following sample application:
`` rosrun metaroom_xml_parser load_labelled_data /path/to/data WayPoint10 ``
This loads all the observations along with the labelled data from the path specified and displays each observation along with the corresponding labelled objects in a visualizer window.
image4 image5 image6

Observation (red) with labelled objects (RGB)

Labelling

For more information on the labelling tool used, please refer to this page.

Download

This dataset is available for download in a single archive file (~ 6 GB). As an alternative, the individual folders and files can be obtained from here, and would have to be downloaded manually.