Face to profile face Dataset

我们提出了一种称为最大空间扰动一致性 (MSPC) 的通用正则化技术，该技术将空间扰动函数 (T) 和平移算子 (G) 强制为可交换 (即T ﹐ G = G ﹖ T)。此外，我们引入了两个用于学习空间扰动函数的对抗性训练组件。第一个让T与G竞争以实现最大扰动。第二个让G和T与鉴别器竞争，以对齐由对象大小变化，对象失真，背景中断等引起的空间变化。在大多数I2I基准测试中，我们的方法优于最先进的方法。我们还引入了一个新的基准，即正面到轮廓面数据集，以强调I2I在实际应用中的潜在挑战。最后，我们进行了消融实验，以研究我们的方法对空间扰动严重程度的敏感性及其对分布对齐的有效性。

We propose a general regularization technique termed Maximum Spatial Perturbation Consistency (MSPC), which enforces the spatial perturbation function (T) and the translation operator (G) to be commutative (i.e., T ∘ G = G ∘ T). Furthermore, we introduce two adversarial training components for learning the spatial perturbation function. The first component pits T against G to pursue maximum perturbation. The second component pits G and T against the discriminator to align the spatial variations induced by object size changes, object distortions, background interruptions, and other similar factors. On most Image-to-Image (I2I) benchmark datasets, our method outperforms state-of-the-art approaches. We also introduce a new benchmark, the Front-to-Profile Face Dataset, to highlight the potential challenges of I2I in real-world applications. Finally, we conduct ablation studies to investigate the sensitivity of our method to the severity of spatial perturbations and its effectiveness on distribution alignment.

搜集汇总

背景与挑战

背景概述

该数据集是一个正面到轮廓面图像转换的基准，用于强调图像到图像转换在实际应用中的潜在挑战。它基于最大空间扰动一致性（MSPC）技术的研究，旨在提升模型在对象大小变化、失真等空间扰动下的性能。