Deep Learning Inference Engine Speed-Accuracy Trade-Off Analysis For Person Search

Loke, Wai-Eu (2019) Deep Learning Inference Engine Speed-Accuracy Trade-Off Analysis For Person Search. Project Report. Universiti Teknikal Malaysia Melaka, Melaka, Malaysia. (Submitted)

Text (24 Pages) Deep Learning Inference Engine Speed-Accuracy Trade-Off Analysis For Person Search - 24 Pages.pdf - Submitted Version Download (16MB)

Abstract

Intel recently released a toolkit dubbed Intel OpenVINO which claims to be able to accelerate deep learning inferencing using only an Intel CPU. A fully functional open-world person search system is developed by using Intel OpenVINO’s inference engine and neural network model as the backend in an attempt to address one of the major challenges of person re-identification. The system is easily scalable in existing camera networks in which only software installations are needed for deployment. Market-1501 large person dataset is used to test the performance of the neural network model. The accuracy of Intel’s pre-trained person re-identification model is relatively accurate when compared with previous works. Furthermore, inferencing speed achieves real-time performance using either CPU or GPU. While there is a definite trade-off in the speed and accuracy between models of different complexity (higher accuracy, lower FPS), there is no trade-off between speed and accuracy when testing between FP16 and FP32 on GPU. As a small contribution, an experiment is carried out to determine the best query image view angle for person reidentification. Improvements for future studies include dynamic feature vectors, dynamic image gallery and using open-world metrics for open-world person reidentification evaluation.

Actions (login required)

View Item