SynRES

该数据集通过密集配对的组合合成训练数据生成自动化管道，这些数据由密集的图像描述驱动合成，可靠的语义对齐机制，以及领域感知的增强技术支持。此外，采用SynRES方法训练的模型在WildRES基准测试中取得了最先进的表现。该数据集包含84,480张合成图像，78,263个表达，以及21,592个遮罩，其任务是指代表达式分割。

This dataset employs an automated pipeline to generate densely paired compositional synthetic training data, which is supported by dense image caption-driven synthesis, reliable semantic alignment mechanisms, and domain-aware augmentation techniques. Furthermore, models trained with the SynRES method achieved state-of-the-art performance on the WildRES benchmark. This dataset contains 84,480 synthetic images, 78,263 referring expressions, and 21,592 masks, targeting the task of referring expression segmentation.