VTBench
收藏资源简介:
VTBench是一个全面的多维基准测试套件,用于评估基于图像的虚拟试穿模型。该数据集由腾讯、复旦大学和厦门大学的研究人员创建,旨在解决现有虚拟试穿模型评估方法的不足。数据集内容包含15个基于不同基础的虚拟试穿模型生成的虚拟试穿图像,以及用于评估这些图像的多种评价指标。数据集创建过程包括收集高质量的测试集、开发可靠的指标以及进行人工偏好标注。数据集的应用领域是虚拟试穿技术,旨在解决现有评估方法无法反映人类感知、测试集仅限于室内场景、缺乏对真实世界场景的复杂性评估等问题。
VTBench is a comprehensive multi-dimensional benchmark suite for evaluating image-based virtual try-on models. Developed by researchers from Tencent, Fudan University, and Xiamen University, this dataset was constructed to address the shortcomings of existing evaluation methods for virtual try-on models. It encompasses virtual try-on images generated by 15 distinct virtual try-on models based on different foundational frameworks, alongside multiple evaluation metrics for assessing these generated images. The dataset creation workflow includes collecting high-quality test datasets, developing reliable evaluation metrics, and carrying out human preference annotation tasks. Targeted at the virtual try-on technology domain, VTBench aims to resolve several key limitations of current evaluation practices: their inability to reflect human perceptual preferences, the restriction of test datasets solely to indoor scenes, and the absence of evaluations for the complexity of real-world scenarios.




