AVA-BENCH
收藏资源简介:
AVA-BENCH是一个针对视觉基础模型(VFMs)的评估基准,旨在明确区分14种基本的视觉能力(AVAs),这些能力是解决复杂视觉推理任务的基础技能,如定位、深度估计和空间理解。该数据集通过将AVAs解耦并匹配每个AVAs的训练和测试分布,能够准确指出VFM在哪些方面表现优异或不足。AVA-BENCH涵盖了广泛的应用领域,旨在为下一代VFMs的发展奠定基础。
AVA-BENCH is an evaluation benchmark for visual foundation models (VFMs). It is designed to clearly disentangle and distinguish 14 fundamental visual abilities (AVAs), which are core skills for solving complex visual reasoning tasks including localization, depth estimation, and spatial understanding. By disentangling the AVAs and aligning the training and test distributions corresponding to each AVA, this benchmark can accurately pinpoint where VFMs excel or underperform. AVA-BENCH covers a wide range of application domains, aiming to lay a solid foundation for the development of next-generation visual foundation models.




