Master Thesis Defense on Video Conferencing Quality at DTU Fotonik

Video conferencing has become an essential tool for hybrid meetings, where some participants join physically at a shared location while others collaborate remotely using conferencing technology. This has driven the development of various video capture devices, including traditional laptop-mounted webcams and more advanced video conferencing systems. However, a variety of sources pose challenges to video quality transmitted during such meetings, with problem sources ranging from "external circumstances" such as compression and transmission artifacts to artifacts stemming from the Image signal processing (ISP) pipeline called "device induced distortions", which if not properly treated, can introduce visual degradations such as noise, blur, or color distortions.

While the primary goal of this thesis is to understand the perceptual video quality in video-conferencing scenarios, the preliminary research suggests the need for attributes that support evaluation of device-induced distortions such as availability of relevant test material and an efficient design of experiment to conduct subjective test. Thus we begin with sourcing relevant test material which consists of 8 video-conferencing samples with diverse room sizes, lighting conditions, and participant actions. On top of these samples, we then apply different levels of synthetic device-induced distortions which are - shot noise, optical blur, and temporal jitter. These distortions are introduced simultaneously to study their combined effect on video quality, which led to generation of 192 videos out of 8 samples.

To evaluate the impact of these degradations, we compute a range of full-reference video quality metrics, namely PSNR, SSIM, MS-SSIM, VMAF, and ColorVideoVDP (CVVDP) and subsequently correlated with subjective scores. The subjective study was conducted across 21 assessors at GN multimedia laboratory under controlled and well calibrated environment using SenseLabOnline platform. The Mean Opinion scores (MOS) were obtained across all the assessors. Finally, we analyze the correlation between objective scores and subjective judgments to identify which metric aligns accurately with perceptual response. Our findings reveal that CVVDP achieves the highest correlation to the MOS score with SROCC of 0.784 and PLCC of 0.823.