DJI-day, Bosonplus-day, Bosonplus-night
收藏资源简介:
本文介绍了ThermalGen,一个自适应的基于流的生成模型,用于RGB到热成像的图像转换。为了支持大规模训练,我们整理了八个公开的卫星-航空、航空和地面RGB-T配对数据集,并引入了三个新的卫星-航空RGB-T数据集,包括DJI-day、Bosonplus-day和Bosonplus-night。这些数据集在不同时间、传感器类型和地理区域捕获,包含了丰富的RGB和热成像数据。ThermalGen通过利用RGB图像和特定于数据集的风格嵌入,实现了在不同视角、传感器特性和环境条件下生成高质量的热成像。这些数据集的发布为视觉-热传感器融合和跨模态任务提供了重要的数据支持,有助于推动相关领域的研究和应用。
This paper introduces ThermalGen, an adaptive flow-based generative model for RGB-to-thermal image translation. To support large-scale training, we curated eight publicly available paired RGB-T datasets from satellite-borne aerial, aerial and ground-based scenarios, and introduced three new satellite-borne aerial RGB-T datasets, namely DJI-day, Bosonplus-day, and Bosonplus-night. These datasets are captured across different time periods, sensor types and geographic regions, containing abundant RGB and thermal imaging data. ThermalGen generates high-quality thermal images across diverse viewpoints, sensor characteristics and environmental conditions by leveraging RGB images and dataset-specific style embeddings. The release of these datasets provides critical data support for visual-thermal sensor fusion and cross-modal tasks, which facilitates the advancement of research and applications in related fields.

- 1通过纽约大学 · 2025年



