Datasets

This page is a collection of high-quality datasets produced by VENTURI partners.

Some datasets are directly related to specific algorithms designed and implemented during the project, especially in the context of WP4. They contain all relevant data for the considered algorithm; i.e. any combination of synchronized audio and/or video and/or hardware sensor (accelerometers, gyroscopes, compass, GPS, pressure sensor, etc.) streams. In some cases, high-quality ground truth data is provided to better evaluate algorithm performances. When possible, links to well known public external datasets are provided as well.

Other datasets presented here are representative sequences of the real world Use Cases addressed by VENTURI, as defined in WP2. Also in this case, synchronized streams and ground truth data are provided, when possible.

Some of the datasets are shared with the community in the public section of this website (Public Datasets). This page lists all datasets, including those sequences that are to be considered private to the VENTURI consortium.

Use Case Datasets
Title Video Audio Sensors GT Device(s) Visibility Notes
Use Case Year 1 (gaming) preliminary sequences, with markers Yes No Yes No Galaxy Nexus consortium
UCY2 datasets (tbd)
UCY3 datasets (tbd)
Algorithm Specific Datasets
Title Video Audio Sensors GT Device(s) Visibility Notes
T4.1 Dataset: user activity recognition and map aided positioning No No Yes Yes STE9500, Nexus S consortium T4.1
Mountain Dataset: outdoor video sequences with GT orientation Yes No Yes Yes Sony Ericsson XPERIA Arc S public T4.3.2
Building Dataset: collection of pictures with GT pose Yes No No Yes iPhone4 public T4.3, T4.4: both Building datasets share the same reference system
Building Dataset: xyzRGB laser-scanner points Yes No No Yes Laser Scanner consortium T4.4: both Building datasets share the same reference system
Fifteen Scene Categories, Ponce Research Group Yes No No Yes external External dataset (link), for testing scene classification algorithms in T4.3.1
ICDAR 2011: Reading Text in Scene Images Yes No No Yes external External dataset (link), for testing text detection and recognition algorithms in T4.3.3
KAIST Scene Text Database Yes No No Yes external External dataset (link), for testing text detection and recognition algorithms in T4.3.3
 The Street View Text Dataset Yes No No Yes external External dataset (link), for testing text detection and recognition algorithms in T4.3.3
KITTI Vision Benchmark Suite Yes No Yes Yes external External dataset (link), for testing optical flow algorithms in T4.4

 

Use Case Year 1 (gaming) preliminary sequences, with markers

The sequences have been captured on a Galaxy Nexus (ICS) using the data acquisition tool available on SteerForge. This dataset was acquired by metaio to stir the discussion about the content of the final dataset for the gaming use case. The dataset contains five sequences: each sequence has one or more videos along with the sensors data dump. The sequences are:

  1. Roundabout: long and full circle around the models with always a marker in the field of view, multiple markers are placed close to the city model
  2. Closeup: close view of two or three places in the virtual city, starting from far away, get closer, almost to street level, and back.
  3. From atop: for the same places  in 2,  a medium distance sequence centered on the first, than move to the second, etc.
  4. On the streets: one road with a 90 degrees turn, followed  closely with the camera to simulate a user driving a virtual car.
  5. Gaming: a longer sequence simulating a gaming session

The sequence is available for download here (256M, see this page for login details).

 

T4.1 Dataset: user activity recognition and map aided positioning

All datasets related to task T4.1 were acquired by INRIA and can be downloaded here (zip, 41.0M, created on Mar 28, 2012) in a single zip file. A document describing the dataset can be downloaded here (pdf, 90K).

We have used two platforms to capture the data: the VENTURI platform and a Samsung Nexus S device. For each dataset we precise the device position with respect to the user (in-hand, chest-mounted, belt-mounted, or in-the-pocket), geometry of the path, distance walked (if needed) and folder name (inside the zip file) where data can be found.

T4.1 Dataset includes:

  • Step detection and estimation of the distance walked (for T4.1.1). Geometry of the path: 30m straight line. GT: two markers at the start and the end points. User height for the acquisition: 165cm. Distance walked: 30m.
  • Stairs detection (for T4.1.2). Geometry of the path: 8 steps, straight line (going down and going up). GT: none.
  • Running detection (for T4.1.2). Geometry of the path: 30m straight line. GT: none.
  • Sitting detection (for T4.1.2).
  • Jumping detection (for T4.1.2).
  • First map aided positioning dataset (for T4.1.3). Geometry of the path: 51.5m x 18m. GT: osm file and map provided.
  • Second map aided positioning dataset (for T4.1.3). Geometry of the path: two rectangles (25.5m x 18m). GT: osm file and map provided.
  • Third map aided positioning dataset (for T4.1.3). Geometry of the path: straight lines, curve and 90° turns. GT: osm file, map and YouTube video.

 

Building Dataset: full version with xyzRGB laser-scanner points

The Building Dataset that can be found in the public section of the website (link) includes 20 images of a building and their partial GT pose. In the same location, metaio also performed a laser scanning, producing clouds of xyzRGB points. The reference system of the xyzRGB points is exactly the same used for defining the image poses.

The full version of the Building dataset can be downloaded here (168M, see this page for login details), and includes:

  • The same 20 pictures of the public one.
  • Camera intrinsic parameters.
  • Position and partial orientation of each image.
  • Two laser-scanner point clouds, in VRML and textual xyzRGB format, for a total of about 6.8 million points.

If you want to open and elaborate the VRML point clouds files, you can try MeshLab, a powerful 3D mesh processing software, distributed as free software.

Please note that the laser-scanner point clouds are NOT public and must be used only in the context of the VENTURI consortium.

  1. No comments yet.
  1. No trackbacks yet.
You must be logged in to post a comment.