Embedding-Driven Data Distillation for 360-Degree IQA With Residual-Aware Refinement

TL;DR

This work tackles data inefficiency in 360-degree IQA by introducing an embedding similarity-based post-sampling patch refinement that selects a compact, informative subset from an initially collected patch pool. The method casts patch embeddings into a low-dimensional space while preserving intrinsic similarity and employs residual analysis to prune redundant patches, yielding a universal, model-agnostic preprocessing step. Empirical results on CVIQ, OIQA, and MVAQD show that 40-50% of patches suffice to match or exceed baselines, with 20-40% reductions in computation when integrated with diverse SOTA architectures. The approach demonstrates strong robustness to distance metrics and embedding dimensionality, and generalizes across CNN- and transformer-based IQA models, offering a practical path to scalable, data-efficient 360-degree quality assessment.

Abstract

This article identifies and addresses a fundamental bottleneck in data-driven 360-degree image quality assessment (IQA): the lack of intelligent, sample-level data selection. Hence, we propose a novel framework that introduces a critical refinement step between patches sampling and model training. The core of our contribution is an embedding similarity-based selection algorithm that distills an initial, potentially redundant set of patches into a compact, maximally informative subset. This is formulated as a regularized optimization problem that preserves intrinsic perceptual relationships in a low-dimensional space, using residual analysis to explicitly filter out irrelevant or redundant samples. Extensive experiments on three benchmark datasets (CVIQ, OIQA, MVAQD) demonstrate that our selection enables a baseline model to match or exceed the performance of using all sampled data while keeping only 40-50% of patches. Particularly, we demonstrate the universal applicability of our approach by integrating it with several state-of-the-art IQA models, incleasy to deploy. Most significantly, its value as a generic,uding CNN- and transformer-based architectures, consistently enabling them to maintain or improve performance with 20-40\% reduced computational load. This work establishes that adaptive, post-sampling data refinement is a powerful and widely applicable strategy for achieving efficient and robust 360-degree IQA.

Figures (7)

• Figure 1: Overview of the proposed embedding similarity-based selection framework. The algorithm transforms patch embeddings to a low-dimensional space while preserving similarity structure, then uses residual analysis to identify and select the most informative patches.
• Figure 2: Minimum selection rates required to exceed baseline performance across different sampling strategies. Radial distance from center indicates efficiency (farther = more efficient, requiring less data).
• Figure 3: Performance trends of PLCC and SRCC metrics as a function of the h-parameter across three datasets (CVIQ: left, OIQA: midle, MVAQD: right). The scatter points represent individual performance measurements, while solid and dashed lines show mean trends with confidence intervals ($\pm$1 standard deviation).
• Figure 4: Statistical significance of various factors (y-axis) on the performances: (top) PLCC (bottom) SRCC, across the used datasets (x-axis).
• Figure 5: Convergence behavior of the objective function (log-scale) across distance metrics and a subset of embedding dimensionalities ($h$).

Theorems & Definitions (6)

• Lemma 1
• proof
• Lemma 2
• proof
• Theorem 1
• proof