Survey on the State-Of-The-Art Methods for Objective Video Quality Assessment in Recognition Tasks
This paper is a technical report, presenting a survey on the state-of-the-art methods for objective video quality assessment in recognition tasks. It bases on the most up-to-date solutions, developed by various research teams. The study considers, among o
- PDF / 1,708,694 Bytes
- 19 Pages / 439.37 x 666.142 pts Page_size
- 12 Downloads / 152 Views
AGH University of Science and Technology, 30059 Krak´ ow, Poland [email protected], [email protected] Huawei Technologies Duesseldorf GmbH, 40549 Duesseldorf, Germany [email protected] http://vq.kt.agh.edu.pl
Abstract. This paper is a technical report, presenting a survey on the state-of-the-art methods for objective video quality assessment in recognition tasks. It bases on the most up-to-date solutions, developed by various research teams. The study considers, among others, solutions developed by the AGH University research team, including the contributions to ITU-T Recommendation P.912 (dealing with video quality assessment methods for recognition tasks) as well as the video quality indicators (available at http://vq.kt.agh.edu.pl/). In particular, we consider evaluation metrics based on a trade-off between computer vision performance and compression efficiency. Keywords: Closed-Circuit Television (CCTV) · Video surveillance · Advanced Driver Assistance System (ADAS) · Self-driving systems · Robot-based industrial production · Quality of Experience (QoE) · Quality of Service (QoS) · Metrics · Evaluation · Performance · Target Recognition Video (TRV) · Computer Vision (CV) · Video Quality Indicators (VQI) · Key Performance Indicators (KPI).
1
Problem Introduction
Nowadays, we have many metrics for overall Quality of Experience (QoE), both Full-Reference ones, like Peak Signal–to–Noise Ratio – PSNR or Structural Similarity – SSIM and No–Reference ones, like video quality indicators, successfully used in video processing systems for video quality evaluation. However, they are not appropriate for recognition tasks analytics (Target Recognition Video, TRV). Given the use of TRV, qualitative tests do not focus on the subject’s satisfaction with the video sequence quality, but instead, they measure how the subject uses TRV to accomplish certain tasks [14]. Purposes of this may include: Supported by Huawei Innovation Research Program (HIRP). c Springer Nature Switzerland AG 2020 A. Dziech et al. (Eds.): MCSS 2020, CCIS 1284, pp. 332–350, 2020. https://doi.org/10.1007/978-3-030-59000-0_25
Survey on the State-Of-The-Art Methods for Objective Video Quality
– – – – – –
333
video surveillance – recognition of vehicle license plate numbers, telemedicine/remote diagnostics – correct diagnosis, fire safety – fire detection, rear backup cameras – parking the car, games – spotting and correctly reacting to a virtual enemy, video newscasts and reports editing – video summarization [6,7].
Since the number of for example surveillance cameras is growing extremely fast, the likelihood that automatic systems will be used to carry out this task is increasing as well. As presented in many scientific publications, automatic recognition algorithms are the clue to handle recognition tasks. Concerning the entertainment video, there were performed researches into the content parameters that most affect perceptual quality. These parameters form a framework creating predictors, and thus developing objective measurements, through the
Data Loading...
