PaperlayerView exact revision on arXiv

Source-linked AI summary

Reduced Reference Perceptual Quality Model and Application to Rate Control for 3D Point Cloud Compression

Qi Liu, Hui Yuan, Raouf Hamzaoui, Honglei Su, Junhui Hou, Huan Yang

arXiv:2011.12688v1eess.IVcs.CV

TL;DR

The paper addresses the lack of a low-cost perceptual quality measure for 3D point clouds that correlates with subjective quality. It proposes a reduced-reference linear model for V-PCC based on quantization parameters and two original-content features, achieving high correlation with MOS and improving perceptual rate control at the same bitrate.

  • Problem

    Existing 3D point-cloud quality measures either have high computational complexity or do not accurately predict visual quality under coding distortion.

  • Method

    The paper develops a reduced-reference linear quality model using V-PCC geometry and color quantization parameters, with coefficients estimated from CFGD and CBMV content features.

  • Results

    PLCC and SRCC between predicted and actual MOS were both 0.91, and the proposed rate-control algorithm achieved much better subjective quality than P2PES.

  • Takeaways & Limitations

    The model supports perceptually optimized V-PCC rate control for static 3D point clouds at a given target bitrate.

Abstract

from arXiv · show

In rate-distortion optimization, the encoder settings are determined by maximizing a reconstruction quality measure subject to a constraint on the bit rate. One of the main challenges of this approach is to define a quality measure that can be computed with low computational cost and which correlates well with perceptual quality. While several quality measures that fulfil these two criteria have been developed for images and video, no such one exists for 3D point clouds. We address this limitation for the video-based point cloud compression (V-PCC) standard by proposing a linear perceptual quality model whose variables are the V-PCC geometry and color quantization parameters and whose coefficients can easily be computed from two features extracted from the original 3D point cloud. Subjective quality tests with 400 compressed 3D point clouds show that the proposed model correlates well with the mean opinion score, outperforming state-of-the-art full reference objective measures in terms of Spearman rank-order and Pearsons linear correlation coefficient. Moreover, we show that for the same target bit rate, ratedistortion optimization based on the proposed model offers higher perceptual quality than rate-distortion optimization based on exhaustive search with a point-to-point objective quality metric.

I. INTRODUCTION

3D point clouds require efficient compression because geometry and color distortions affect reconstructed perceptual quality. The paper proposes a reduced-reference model for V-PCC and applies it to perceptually optimized rate control.

  • 3D point clouds can contain millions or billions of points, creating substantial storage and transmission demands.
  • V-PCC compresses both geometry and color, whose distortions influence the perceived quality of reconstructed 3D point clouds.
  • Existing full-reference metrics are computationally demanding or fail to predict visual quality accurately, especially under coding distortion.
  • The proposed reduced-reference model predicts V-PCC perceptual quality from geometry and color quantization steps and supports rate-distortion optimized rate control.
  • Prior subjective datasets often provided limited reference content, few assessed point clouds, or did not focus specifically on compression degradation.

III. SUBJECTIVE QUALITY ASSESSMENT

The subjective test dataset evaluates static point-cloud compression across many high-quality reference objects and encoding degradation levels. Stimuli were generated from multiple viewpoints to support perceptual assessment.

  • The study assembled sixteen high-quality reference point clouds, including objects such as Bag, Banana, Biscuits, Cake, and Cauliflower.
  • The dataset was designed to address the need for sufficient reference content and varied encoding degradation levels in subjective testing.
  • Pictures were generated from 360 viewpoints for each point cloud to represent the objects from horizontal and vertical directions.
  • The subjective evaluation recruited 30 subjects, comprising 15 males and 15 females aged 20 to 35 with normal or corrected-to-normal vision.

B. Subjective evaluation

Observers rated compressed point-cloud videos using a double-stimulus impairment procedure with a continuous 100-point scale. The evaluation compared original and distorted videos shown simultaneously.

  • The study used the Double-Stimulus Impairment Scale methodology with a continuous 100-point rating scale.
  • Original and distorted videos generated from each 3D point cloud were displayed side by side during evaluation.
  • Observers completed a training session using different point clouds to learn distortion types and quality levels without becoming familiar with evaluation content.
  • The two-hour test for each subject was divided into four sections with three five-minute breaks to reduce fatigue effects.

C. Data post-processing

Subjective ratings were standardized per observer, screened for outliers, and averaged across valid subjects to obtain MOS values. Individual-observer agreement with MOS was also assessed.

  • Raw ratings were converted to Z-scores using each observer’s rating mean and standard deviation.
  • Outlier ratings were discarded from each participant, while no participants were removed from the evaluation.
  • The MOS for each degraded point cloud was calculated as the average Z-score across all valid subjects.
  • Mean PLCC and SRCC between individual observer scores and MOS were both 0.84, indicating substantial agreement among subjects.

IV. PROPOSED QUALITY METRIC MODEL

The proposed model relates perceptual distortion to geometry and color quantization steps through progressively simplified linear relationships. Statistical analysis removes the interaction term, yielding a convex model with strong fitting accuracy.

  • The initial model represents MOSc using geometry and color quantization terms, their interaction, and an intercept.
  • Removing the interaction term makes the quality model convex, supporting applications such as rate-distortion optimization.
  • The interaction Qg·Qc has much lower statistical significance than Qg and Qc in the two-way ANOVA.
  • The simplified model achieves an SCC up to 0.949, with an average SCC of 0.914 across evaluated 3D point clouds.

V. MODEL PARAMETER PREDICTION USING CONTENT FEATURES

Content complexity affects perceived distortion at identical quantization steps, so the model parameters are content dependent. The paper therefore predicts them from local and global color-texture features.

  • At identical quantization steps, richly textured content such as Cake has lower MOSc and higher MOS than simply textured content such as Ping-pong bat.
  • The model parameters are highly content dependent because content has a concealing effect on coding distortion.
  • The proposed parameter-prediction features are CFGD, representing local color fluctuation over geometric distance, and CBMV, representing global color block mean variance.

A. Color fluctuation over geometric distance (CFGD)

CFGD is a local texture feature that captures color fluctuation over geometric distance in a 3D point cloud. It uses neighboring color differences, weighted by point distance, and aggregates these values across all points.

  • A. Color fluctuation over geometric distance (CFGD): CFGD describes local content characteristics by measuring color gradient variation over geometric distance.The feature uses neighboring color intensity differences around each point.
  • A. Color fluctuation over geometric distance (CFGD): For each point, CFGD averages neighboring color-intensity differences using distances to the K nearest neighbors.The paper considers only the Y luminance component for simplicity.
  • A. Color fluctuation over geometric distance (CFGD): The point-level CFGD values are aggregated over all T points to obtain the 3DPC-level feature.

B. Color block mean variance (CBMV)

CBMV is a global feature constructed from voxelized 3D point clouds by measuring color variation within non-empty voxels. It uses within-voxel color deviations to characterize global content.

  • B. Color block mean variance (CBMV): CBMV is a global feature built from the color variation of points inside non-empty voxels.The point cloud is assumed to be voxelized before CBMV computation.
  • B. Color block mean variance (CBMV): For each non-empty voxel, CBMV uses the number of points, point colors, and the voxel’s mean color value.B denotes non-empty voxels, D the points in each voxel, C(pij) point color, and µi the voxel color mean.

C. Model parameter estimation

The model estimates three perceptual-quality parameters from two extracted point-cloud features using a generalized linear model. Training content determines the coefficient matrix, which is then used to predict MOS and compare the model with full-reference metrics.

  • C. Model parameter estimation: A generalized linear model predicts each perceptual-quality parameter from two extracted features for every 3D point cloud.The two features are represented in the feature vector used by the predictor.
  • C. Model parameter estimation: The predictor estimates the three-parameter vector ˆPm from FmH, where Fm contains a constant and two feature values.The coefficient matrix H is optimized to minimize prediction error.
  • C. Model parameter estimation: Eight 3DPCs train H, while eight different 3DPCs form the test set; voxel size 643 is used for CBMV.The training and testing objects cover different content characteristics.
  • C. Model parameter estimation: Using Hopt and extracted features, the method predicts model parameters and then obtains MOS for evaluation against actual MOS.PLCC, SRCC, and RMSE are used to assess prediction accuracy.
  • C. Model parameter estimation: 0.9133 PLCC for the proposed reduced-reference metric exceeded the 0.4027–0.8199 PLCC range of the compared full-reference metrics.The proposed metric also achieved higher SRCC and lower RMSE than the compared metrics.

VI. APPLICATION

The proposed perceptual model is embedded in V-PCC rate control to select geometry and color quantization parameters under a target bitrate. Compared with point-to-point exhaustive search, it achieves better rate-MOS performance with lower complexity.

  • VI. APPLICATION: The rate-control objective selects geometry and color QPs that maximize perceptual quality under an overall target bitrate.Geometry and color bitrates must satisfy the target-rate constraint.
  • VI. APPLICATION: The resulting optimization uses the perceptual model p1Qg + p2Qc + p3 together with geometry and color rate models.The parameters p1, p2, and p3 describe the perceptual-quality model.
  • VI. APPLICATION: The encoder extracts CFGD and CBMV, applies pretrained H, and solves for optimal Qg,opt and Qc,opt using an optimization method.An interior-point method or another convex optimization method can solve the problem.
  • VI. APPLICATION: The proposed algorithm achieves better rate-MOS performance than P2PES with much lower complexity.The comparison uses rate-MOS curves across point clouds with target bitrates listed in Table VII.
  • VI. APPLICATION: The proposed RR model-based rate control achieves significant subjective-quality improvement over P2PES in the visual comparisons.Figure 14 compares original, proposed, and P2PES reconstructions at specified target bitrates.

VII. CONCLUSION

The proposed reduced-reference linear model predicts perceptual quality from V-PCC geometry and color quantization parameters using two reference-point-cloud features. It achieves high correlation with subjective scores and supports rate control with better subjective quality than P2PES, while evaluation on newer MPEG PCC content remains future work.

  • The reduced-reference linear model predicts V-PCC compressed point-cloud quality from geometry and color quantization parameters using two features of the reference point cloud.Its three coefficients are estimated from a training set of reference point clouds and the CFGD and CBMV features computed from the test reference point cloud.
  • 0.91 PLCC and 0.91 SRCC were obtained between predicted and actual MOS values in subjective experiments on static 3D point clouds.The experiments used high-quality point clouds selected from the WPC dataset because the MPEG PCC collection was limited.
  • The optimized rate-control algorithm achieved much better subjective quality than P2PES.The algorithm was proposed to demonstrate an application of the reduced-reference quality model to 3D point-cloud compression.
  • Evaluation on the high-quality 3D point clouds recently provided by the MPEG PCC group remains future work.
Loading 2011.12688v1…