Publication Date

Fall 2021

Degree Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Engineering

Advisor

Mahima A. Suresh

Subject Areas

Computer engineering

Abstract

Depth estimation algorithms are useful components of computer vision systems toassess video streams on urban streets. They can provide important information about the street space and improve situational awareness for humans. Deep learning algorithms for depth estimation are slow in providing on-time street evaluation for the users. To provide situational awareness to humans, the inference time needs to be small, so that results are fresh and meaningful. This paper explores approaches like switching to efficient convolutional neural network, quantization and pruning to improve inference time with a little compromise on the performance. We explore the impact of each of these methods independently and introduce a hybrid method. We evaluate the execution time, resource utilization, and performance of various state-of-the-art depth estimation algorithms. We compare these with our approach of using the three optimization techniques both independently and in hybrid. We observe that using these optimization techniques can improve the inference time dramatically, with a 57.3% speedup in inference time and a 94.5% reduction in memory utilization while improving the object level performance (RMSE) by 3.8%.

Download

Share

COinS