SAVOIAS
收藏资源简介：
我们介绍了 SVOIAS 一个视觉复杂性数据集，该数据集包含来自场景、广告、可视化和信息图表、对象、室内设计、艺术和至上主义七个图像类别的 1,400 多张图像。每个类别中的图像描绘了不同的特征，包括各种低级和高级特征、对象、背景、纹理和图案、文本和图形。 SVOIAS 的基本事实是通过使用强制选择方法众包 37,000 多个成对的图像比较以及使用 Figure-8 众包平台的 1,600 多个贡献者来获得的。然后使用 Bradley-Terry 方法和矩阵完成将得到的相对分数转换为绝对视觉复杂度分数。在应用五种最先进的算法来分析 SVOIAS 数据集中图像的视觉复杂性时，我们发现从这些基线工具获得的分数仅与至上主义类别中抽象模式的众包标签具有良好的相关性。皮尔逊相关 r = 0.84）。对于其他类别，特别是对象和广告类别，相关系数较低（分别为 r = 0.3 和 0.56）。这些发现表明（1）最先进的方法大多不足，（2）SVOIAS 支持特定类别的方法开发，这可能会改善视觉复杂性分析对特定应用领域（包括计算机视觉）的影响。
We introduce SVOIAS, a visual complexity dataset containing over 1,400 images across seven image categories: scenes, advertisements, visualizations and infographics, objects, interior designs, artworks, and suprematism. Images in each category depict diverse features, including various low-level and high-level features, objects, backgrounds, textures and patterns, text and graphics. The ground truth of SVOIAS was obtained by crowdsourcing over 37,000 pairwise image comparisons using a forced-choice method, with more than 1,600 contributors via the Figure-8 crowdsourcing platform. Subsequently, the obtained relative scores were converted into absolute visual complexity scores using the Bradley-Terry method and matrix completion. We applied five state-of-the-art algorithms to analyze the visual complexity of images in the SVOIAS dataset, and found that the scores obtained from these baseline tools only correlated well with the crowdsourced labels of abstract patterns in the suprematism category (Pearson correlation r = 0.84). For other categories, especially the object and advertisement categories, the correlation coefficients were lower (r = 0.3 and 0.56, respectively). These findings indicate that (1) most state-of-the-art methods are insufficient, and (2) SVOIAS supports the development of category-specific methods, which may improve the applicability of visual complexity analysis for specific application domains including computer vision.
