23
COLOR-PLUS-DEPTH LEVEL-OF-DETAIL IN 3D TELE-IMMERSIVE VIDEO: A PSYCHOPHYSICAL APPROACH Wanwin Wu et la., MONET Group UIUC ACM MM 2011 Best Paper

COLOR-PLUS-DEPTH LEVEL-OF- DETAIL IN 3D TELE-IMMERSIVE VIDEO: A PSYCHOPHYSICAL APPROACH Wanwin Wu et la., MONET Group UIUC ACM MM 2011 Best Paper

Embed Size (px)

Citation preview

Color-plus-Depth Level-of-Detail in 3D Tele-immersive Video: A Psychophysical Approach

Color-plus-Depth Level-of-Detail in 3D Tele-immersive Video: A Psychophysical ApproachWanwin Wu et la., MONET Group UIUCACM MM 2011 Best PaperTele-immersion

Psychophysics (PP)

Psychophysical Thresholds

Problems

Real-time 3D model constructionMulti-party communicationInter/Intra-node synchronizationComputational resource-hungryNetworking resource-hungry

Computational resource BalancingConstruction of 3D Models

Parameters of Color-plus-Depth Level-of-Detail (CZLoD)Detailing parameter of CZLoD: THvar (Threshold of Variance)- Recursively refining bisection until the variance within every polygon (triangle) is less than THvar - THvar decides the size of triangles- THvar decides both texture (color) and spatial resolutionDegradation of CZLoD: DR (Degradation Ratio)- Fi: the ith 3D frame- Nx(Fi): number of vertices in Fi if THvar = x

GoalsFinding the PP thresholds in tele-immersive videos- JNDG (Just Noticeable DeGradation)- JUADG (Just UnAcceptable DeGradation)Utilize the thresholds on constructing a real-time adaptation scheme for resource balancing- QoS Monitor- Decision Engine- Variance CalculationGoal 1: Finding PP ThresholdsPhysical stimuli: DRPerceptual quantity: JNDG and JUADG(participant was asked if he/she could tell any difference between the two clips, and whether he/she thought any video had unacceptable quality)Experiment method: Ascending Method of Limit

Video

Finding PP Thresholds (conti.)Participants: - 16 graduate students of CS UIUC- all had normal or corrected vision- 4 Indian, 3 American, 2 Chinese, 2 German, 2 Bangladeshi, 1 Mexican, 1 South African- 6 women (37.5%) and 10 men (62.5%)- 5 experts of tele-immersion (31.25%) and 11 novices (68.75%)Mapping From Thvar To DRVery much content-dependent

The input parameter (THvar) is manually adjusted to form a series of gradually degrading videos

ResultsThe average thresholds among participants and among the four videos areJNDG: 61.5%; JUADG: 81.25%Thresholds of low resolution videos are lower (easier for users to notice degradations)Influence of content is less with higher resolutionThe size of gray zone is 10~20%

Goal 2: Adaptation Scheme

QoS MonitorInput: the last 3D frame displayedAnalyze: DR and construction time (frame rate)

Output- DR value - abnormal frame rate (FR) event: if (FR > THh || FR < THl)

More Details On QoS MonitorAccording to [16], (THh, THl) is set as (8, 12)In calculation of DR, N0(Fi) is periodically computed to reduce complexityDR here is hence defined as min{1, DR(Fi)}The frame rates are measured within a sliding window of 5 frames

Decision EngineInput: DR value and abnormal FR reportOutput: target DR and DR error (explained later)The purpose is fairly simple, actually- If FR is too low, increase it (by d)- If FR is too high, decrease it (by u)The nontrivial are the amounts of changes (d and u)The changes can follow several protocols- constant- AIMD- ...

What about the PP Thresholds?Three decision zones

The adaptation zone = [JNDG-Bn, JUADG+Ba]d and u are set as 0 outside the adaptation zone

Variance CalculatorInput: target DR and DR errorOutput: THvarThe same problem again: a mapping F from DR to THvar is neededSolution: least-square regression-based learning

When DR error is larger then some threshold (THerr) then the learning process is triggeredWith ten training points (s0~s9), the median residual is 0.022%

Performance EvaluationFrame rate improvement

(a) w/o extra CPU stress; (b) w/ 16% CPU stress

Performance Evaluation (Conti.)DR THvar mapping

Crowd-sourcing (Youtube) scoring

Summary / Future WorkPP thresholds do exist in tele-immersion PP thresholds can help improving resource balancing between spatial and temporal resolutionsAlthough mapping between parameter setting and actual CZLoD is content dependent and nontrivial, a simple regression-based learning provides sufficient predictionExpect higher computing power to support real-time N0 computation in the futureMore complex protocols to deal with the adjustment (d and u) of FR

RemarksNo variance shown in the experiment results?No statistical support of significance?Sampling of stimuli needs to be non-uniform