You are here: Home Proceedings
Document Actions

Proceedings

Improving Visual Object Detection Using Synthetic Self-Training

Verbesserung der Objekterkennung durch Selbst-Training mit synthetischen Bildern

  1. M.Sc. Christopher Mayershofer Chair of Materials Handling, Material Flow, Logistics Department of Mechanical Engineering Technical University of Munich
  2. B.Sc. Adrian Fischer Chair of Materials Handling, Material Flow, Logistics Department of Mechanical Engineering Technical University of Munich
  3. Prof. Dr.-Ing. Johannes Fottner Chair of Materials Handling, Material Flow, Logistics Department of Mechanical Engineering Technical University of Munich

Abstracts

The current era of supervised learning requires a large corpus of application-specific training data with ground-truth annotations. The creation of large annotated datasets however is a costly endeavor. Moreover, the availability of a large annotated set of training data cannot be guaranteed in certain domains. Self-training attempts to overcome these problems by using a set of labeled data and a potentially infinite pool of unlabeled data to train a model in a semi-supervised manner. Self-training however only works if the annotated data is sufficient for training a strong teacher model, which depending on the application domain, is not always available. In this work, we propose and formulate a simple extension to the self-training paradigm and refer to it as Synthetic Self-Training (SST). SST is able to overcome the aforementioned problem by incorporating synthetically generated images into the training process, therefore improving model performance. Specifically, we address the problem of object detection in a logistics environment and are able to improve the state-of-the-art detection performance on the LOCO dataset by 12% mAP0.5.

Die gegenwärtige Praxis des überwachten Lernens erfordert einen umfangreichen Korpus annotierter Trainingsdaten. Die Erstellung großer annotierter Datensätze ist jedoch ein kostspieliges Unterfangen. Darüber hinaus variiert die Verfügbarkeit eines großen annotierten Trainingsdatensatzes über unterschiedliche Anwendungsbereiche. Selbsttraining versucht, diese Probleme zu überwinden, indem eine Kombination aus annotierten Daten und nicht-annotierten Daten verwendet wird, um ein Modell zu trainieren. Selbsttraining bedarf jedoch einer ausreichenden Menge annotierter Trainingsdaten, um ein starkes Lehrermodell zu trainieren. Diese Arbeit stellt das Synthetische Selbst-Training (SST) vor, eine Erweiterung des konventionellen Selbst-Trainings. SST löst zuvor genannte Problem, durch Einbeziehung synthetisch erzeugte Daten in den Trainingsprozess. Diese Arbeit formuliert SST im Bereich der Visuellen Objekterkennung und zeigt empirische dessen Vorteile. Konkret ermöglicht es SST die Erkennungsgenauigkeit logistikspezifischer Objekte im LOCO Benchmark um 12% mAP0.5 zu verbessern.

Keywords

Fulltext

License

Any party may pass on this Work by electronic means and make it available for download under the terms and conditions of the free Digital Peer Publishing License. The text of the license may be accessed and retrieved at http://www.dipp.nrw.de/lizenzen/dppl/fdppl/f-DPPL_v1_de_11-2004.html.

Number of citations

Visit Google Scholar to find out, how often this paper is cited.