Kinect@home: A crowdsourced RGB-D dataset
2016 (English)In: 13th International Conference on Intelligent Autonomous Systems, IAS 2014, Springer, 2016, Vol. 302, 843-858 p.Conference paper (Refereed)Text
Algorithms for 3D localization, mapping, and reconstruction are getting increasingly mature. It is time to also make the datasets on which they are tested more realistic to reflect the conditions in the homes of real people. Today algorithms are tested on data gathered in the lab or at best in a few places, and almost always by the people that designed the algorithm. In this paper, we present the first RGB-D dataset from the crowd sourced data collection project Kinect@Home and perform an initial analysis of it. The dataset contains 54 recordings with a total of approximately 45 min of RGB-D video. We present a comparison of two different pose estimation methods, the Kinfu algorithm and a key point-based method, to show how this dataset can be used even though it is lacking ground truth. In addition, the analysis highlights the different characteristics and error modes of the two methods and shows how challenging data from the real world is.
Place, publisher, year, edition, pages
Springer, 2016. Vol. 302, 843-858 p.
, Advances in Intelligent Systems and Computing, ISSN 2194-5357 ; 302
Benchmark, Dataset, Reconstruction, RGB-D, SLAM, Benchmarking, Computer programming, Image reconstruction, 3D localization, Data collection, Error mode, Ground truth, Pose estimation, Real-world, Algorithms
IdentifiersURN: urn:nbn:se:kth:diva-181120DOI: 10.1007/978-3-319-08338-4_61ISI: 000377956900061ScopusID: 2-s2.0-84945950370ISBN: 978-3-319-08338-4ISBN: 978-3-319-08337-7OAI: oai:DiVA.org:kth-181120DiVA: diva2:901651
13th International Conference on Intelligent Autonomous Systems, IAS 2014, Padova, Italy, 15 July 2014 through 18 July 2014
QC 201602082016-02-082016-01-292016-07-18Bibliographically approved