TY - GEN
T1 - FENet
T2 - 2020 IEEE International Symposium on Safety, Security, and Rescue Robotics, SSRR 2020
AU - Zhou, Yang
AU - Ge, Rundong
AU - Mcgrath, Gary
AU - Loianno, Giuseppe
N1 - Funding Information:
This work was supported by Qualcomm Research, the ARL grant DCIST CRA W911NF-17-2-0181, and the Technology Innovation Institute. 1The authors are with the New York University, Tandon School of Engineering, Brooklyn, NY 11201, USA. email: {yangzhou, rundong.ge, loiannog}@nyu.edu. 2The author is with Qualcomm Technologies, Inc., 5775 Morehouse Drive, San Diego, USA. email: gmcgrath@qti.qualcomm.com.
Publisher Copyright:
© 2020 IEEE.
PY - 2020/11/4
Y1 - 2020/11/4
N2 - Semantic edge is a geometric-aware semantic feature that can be leveraged in robotic perception systems for increased situational awareness and high-level environment understanding. This sparse representation nicely encapsulates the semantic information (object categories) within the geometric object boundaries. State-of-the-art semantic edge detection approaches require significant computation power and fail to approach real-time performance on embedded devices for robotic applications. In this paper, we present FENet (Fast Realtime Semantic Edge Detection Network), a semantic edge detection approach for robots with Size, Weight, and Power (SWaP) constraints. Specifically, we adopt MobileNetV2 as a lightweight backbone network, and we utilize joint pyramid upsampling to improve the system performance. We further reduce the model complexity and latency through network pruning and multiple upsampling strategies to adapt the model on embedded devices such as NVIDIA Jetson TX2. The proposed method is evaluated on Cityscapes with accuracy performances close to the state-of-the-art methods, but with substantially reduced computational complexity that speeds up the network by a factor of 10. To the best of our knowledge, FENet is the first real-time semantic edge detection network for robotic platforms.
AB - Semantic edge is a geometric-aware semantic feature that can be leveraged in robotic perception systems for increased situational awareness and high-level environment understanding. This sparse representation nicely encapsulates the semantic information (object categories) within the geometric object boundaries. State-of-the-art semantic edge detection approaches require significant computation power and fail to approach real-time performance on embedded devices for robotic applications. In this paper, we present FENet (Fast Realtime Semantic Edge Detection Network), a semantic edge detection approach for robots with Size, Weight, and Power (SWaP) constraints. Specifically, we adopt MobileNetV2 as a lightweight backbone network, and we utilize joint pyramid upsampling to improve the system performance. We further reduce the model complexity and latency through network pruning and multiple upsampling strategies to adapt the model on embedded devices such as NVIDIA Jetson TX2. The proposed method is evaluated on Cityscapes with accuracy performances close to the state-of-the-art methods, but with substantially reduced computational complexity that speeds up the network by a factor of 10. To the best of our knowledge, FENet is the first real-time semantic edge detection network for robotic platforms.
UR - http://www.scopus.com/inward/record.url?scp=85099444153&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85099444153&partnerID=8YFLogxK
U2 - 10.1109/SSRR50563.2020.9292631
DO - 10.1109/SSRR50563.2020.9292631
M3 - Conference contribution
AN - SCOPUS:85099444153
T3 - 2020 IEEE International Symposium on Safety, Security, and Rescue Robotics, SSRR 2020
SP - 246
EP - 251
BT - 2020 IEEE International Symposium on Safety, Security, and Rescue Robotics, SSRR 2020
A2 - Marques, Lino
A2 - Khonji, Majid
A2 - Dias, Jorge
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 4 November 2020 through 6 November 2020
ER -