MultiDensity Rendered (MDR) event optical flow dataset
收藏资源简介:
MDR数据集是由中国电子科技大学和Megvii Technology合作创建的,旨在解决从事件相机中估计光流的问题。该数据集包含80,000个样本,来自53个虚拟场景,每个样本都包含了精确的事件和流标签。数据集的创建过程涉及使用Blender渲染3D场景,包括室内外环境,以及多样化的相机运动,以生成高帧率视频和准确的流标签。MDR数据集不仅适用于光流学习,还能通过其自适应密度模块(ADM)进一步提高性能,该模块能够根据需要调整事件的密度,适用于各种基于事件的任务。
The MDR dataset was co-developed by the University of Electronic Science and Technology of China and Megvii Technology, aiming to address the problem of optical flow estimation from event cameras. This dataset consists of 80,000 samples sourced from 53 virtual scenarios, with each sample containing precise event and optical flow ground-truth labels. The construction of the MDR dataset involves using Blender to render 3D scenes covering both indoor and outdoor environments, as well as diverse camera motions, to generate high-frame-rate videos and accurate optical flow ground-truth labels. The MDR dataset is not only applicable to optical flow learning, but also can further enhance model performance via its Adaptive Density Module (ADM), which can adjust event density as needed and is suitable for various event-based tasks.

- 1Learning Optical Flow from Event Camera with Rendered Dataset中国电子科技大学 · 2023年



