零样本迁移VQA（ZST-VQA）数据集

零样本迁移VQA（ZST-VQA）数据集是由百度研究院创建，旨在模拟人类从输入到输出的知识迁移能力。该数据集通过对现有VQA v1.0数据集进行重组，形成两个任务：零样本答案任务（ZSA）和零样本问题任务（ZSQ）。数据集包含248,349个问题，用于测试模型在未见过的词汇上的迁移能力。创建过程中，数据集从原始训练和测试样本中筛选出共享词汇，并随机抽样形成零样本词汇。该数据集主要应用于视觉问答领域，解决模型在零样本学习上的挑战，检测模型是否能有效处理和迁移新词汇。

The Zero-Shot Transfer Visual Question Answering (ZST-VQA) dataset was developed by Baidu Research, with the goal of simulating the human ability of knowledge transfer from input to output. This dataset is constructed by reorganizing the original VQA v1.0 dataset, and establishes two tasks: the Zero-Shot Answer (ZSA) task and the Zero-Shot Question (ZSQ) task. It contains 248,349 questions, designed to evaluate a model's capacity for knowledge transfer to unseen vocabulary. During the dataset creation process, shared vocabulary is filtered from the original training and test samples, and zero-shot vocabulary is formed via random sampling. This dataset is primarily utilized in the visual question answering domain, to tackle the challenges of zero-shot learning in models, and to verify whether a model can effectively process and transfer novel vocabulary.