Collaborative Inference for Distributed Camera System

Restricted (Penn State Only)
Hakimi, Zeinab
Graduate Program:
Computer Science and Engineering
Master of Science
Document Type:
Master Thesis
Date of Defense:
April 20, 2019
Committee Members:
  • Vijaykrishnan Narayanan, Thesis Advisor
  • John Morgan Sampson, Committee Member
  • Multi View Convolutional Neural Network
  • Distributed System
  • Context Awareness
  • DNN
  • Object Recognition
Recently, it has been veri ed that making use of multiple sources of sensor data can improve the accuracy of inference in distributed networks. However, there are two challenges to achieve this goal: (i ) Sensors inherently provide information with di erent level of quality. Therefore, it is essential to identify the information contribution of each sensor in order to enhance the eciency of the overall system and, (ii ) Unequal information contribution of sensors necessitates to re-examine the assumption of noise tolerance or model failure in distributed systems. To tackle the rst challenge, this thesis proposes a Multi-View Convolutional Neural Network (MVCNN) for distributed camera systems which leverages the likelihood estimation. We used entropy estimation in order to reduce the com- munication cost between front-end and back-end and enhance the performance of object classi cation over the system. Applying our framework to Princeton 3D CAD ModelNet dataset and iLab-80M with 12 views, we reached 89% top-1 classi cation accuracy in tested datasets by pruning 66.7% of sensor nodes. To address the second challenge, we designed a fault-tolerant mechanism for situations when our network can not overcome failures. Experiments demonstrate that our MVCNN maintains nearly-optimal accuracy when the fraction of noisy failures are less than 40%. Moreover, proposed robustness mechanism increases accuracy by 8.54% in the presence of 60% nodes' failures.