Masterarbeit MSTR-2022-31

Tripathy, Bikash: Performance Analysis of Object Detection Models in Context of Mobile Applications.
Universität Stuttgart, Fakultät Informatik, Elektrotechnik und Informationstechnik, Masterarbeit Nr. 31 (2022).
118 Seiten, englisch.

While providing support to the technicians working on the devices, it is quite important that they receive immediate guidance while working on it. One of such solutions is provided using the Augmented Reality solution by TeamViewer using the product ‘TeamViewer Assist AR’ for guiding the user. However, this requires a remote technician to guide the user using an established remote connection.. This work proposed a unique solution by introducing the added intelligence to the device of the user itself to avoid dependencies on the remote connection and remote guidance. This can be achieved by mixing the neural network intelligence along with the augmented reality solutions. In this work, we propose a unique solution to handle this problem using estimating the state of the object in which it is operating using the TensorFlow object detection, and provide useful guidance by placing the 3D markers or objects in the augmented reality world, which would provide an immersive experience to the user. Along with this, edge devices have tight resource constraints, and running neural networks on these devices is quite a computationally intensive task. In addition to this, mobile devices have limited computation capabilities in terms of computational units present such as CPU, GPU, and special vendor hardware accelerators. In this regard, this thesis also analyses the performance of object detection models for on-device machine learning in the context of mobile applications. Our work conducts a detailed study on the family of neural network architecture mixed with augmented reality solutions for Android devices. On top of this, it also looks into performance aspects using the TensorFlow interpreter delegates such as GPU and NNAPI for Android devices. We found the NNAPI delegate performs best among other delegates for TensorFlow. However, we also found lots of issues with model loading using the NNAPI delegate. To the best of our knowledge, this is the first work studying detailed performance analysis for the EfficientDet model architectures mixed with augmented reality solutions.

Abteilung(en)Universität Stuttgart, Institut für Architektur von Anwendungssystemen
BetreuerAiello, Prof. Marco; Bauer, Dennis; Rau, Sascha
Eingabedatum16. September 2022
   Publ. Institut   Publ. Informatik