MD-syn
收藏资源简介:
MD-syn是由华中科技大学和武汉大学联合创建的多模态图像匹配数据集,旨在解决跨模态图像匹配中的数据稀缺问题。该数据集包含4.8亿对图像,涵盖了多种模态(如RGB、红外、深度等)和丰富的场景。数据集通过生成模型从RGB图像扩展而来,继承了RGB数据的匹配标签和多样性。MD-syn的创建过程包括使用数据引擎生成伪模态图像,并确保数据的平衡性和场景覆盖。该数据集广泛应用于跨模态图像匹配任务,显著提升了模型的泛化能力和零样本性能。
MD-syn is a multimodal image matching dataset jointly developed by Huazhong University of Science and Technology and Wuhan University, which aims to address the data scarcity challenge in cross-modal image matching. This dataset contains 480 million image pairs, covering diverse modalities including RGB, infrared, depth, etc., as well as rich scenarios. It is extended from RGB images via generative models, inheriting the matching labels and inherent diversity of RGB data. The development pipeline of MD-syn involves using data engines to generate pseudo-modal images, while ensuring data balance and comprehensive scene coverage. This dataset is widely adopted in cross-modal image matching tasks, and significantly boosts the generalization ability and zero-shot performance of models.

- 1MINIMA: Modality Invariant Image Matching华中科技大学, 武汉大学 · 2024年



