Brown Digital Repository | Item

Full Metadata

Overview

Year:: 2020
Contributor:: Derman, Can Eren (creator); Bahar, Iris (thesis advisor); Taubin, Gabriel (reader); Brown University. School of Engineering (sponsor)
Genre:: theses
Subject:: Neural networks (Computer science); Machine learning; Computer vision; Robotics
DOI: https://doi.org/10.26300/1407-dm15

Files

Description

Abstract:: Advancements in machine learning techniques have encouraged scholars to focus on convolutional neural network (CNN) based solutions for object detection and pose estimation tasks. Most robotic applications fail to achieve desired accuracies in real world conditions because CNNs' performance depends heavily on the similarity between the training and testing data. In their work, Generative Robust Inference and Perception (GRIP), the Bahar Group has introduced a second stage to assess CNN outputs and increase the robustness of the estimations, especially in adversarial conditions. Although the iterative likelihood weighting, introduced by GRIP's second stage, increases the pose estimation accuracy significantly, it is not optimized for runtime. In this thesis, we propose FPGA Accelerated Simultaneous Trigonal Edge Detection and Rasterization (FASTER) as a faster alternative to GRIP. On top of producing robust pose estimations in adversarial scenarios, FASTER optimizes the bottlenecks of the second stage algorithm by creating a highly parallelized custom circuit on an FPGA with a novel edge detection algorithm. We benchmark FASTER by comparing it to GRIP's accuracy on dark and occluded scenes. Our results show that FASTER achieves a significant speedup with outstanding energy savings compared to GRIP. These results are especially impressive given that GRIP is implemented on a much more sophisticated hardware with higher clock frequency.
Notes:: Senior thesis (ScB)--Brown University, 2020; Concentration: Computer Engineering

Content

Access Conditions

Rights: In Copyright
Restrictions on Use: All rights reserved. Collection is open for research.

Citation

Derman, Can Eren, "FASTER: FPGA Accelerated Simultaneous Trigonal Edge Detection and Rasterization for Robust Robotic Manipulation" (2020). Engineering Theses and Dissertations. Brown Digital Repository. Brown University Library. https://doi.org/10.26300/1407-dm15

Relations

Collection:

Engineering Theses and Dissertations

Theses and Dissertations for the Engineering department.
...

FASTER: FPGA Accelerated Simultaneous Trigonal Edge Detection and Rasterization for Robust Robotic Manipulation