Real Scenes with Spike and Flow (RSSF)
收藏资源简介:
Real Scenes with Spike and Flow (RSSF)是一个用于训练和评估基于脉冲的光流的数据集。该数据集基于Slow Flow数据集生成,使用高速摄像机捕捉。训练集包含31个场景,测试集包含10个场景。训练集有三种不同分辨率和脉冲帧数的场景,总计超过9.6k个流场和193k+脉冲帧。评估数据集包含11个场景,每个场景选择前200个流场。为标准化评估数据,使用中心裁剪使每个脉冲帧和流场的宽度为1024,高度超过768的图像高度为768。评估数据集的流场和脉冲帧总数分别为2.2k和44.22k。
The Real Scenes with Spike and Flow (RSSF) dataset is designed for training and evaluating spike-based optical flow algorithms. Derived from the Slow Flow dataset, it is captured using high-speed cameras. The training set comprises 31 scenes, while the test set includes 10 scenes. The training set features scenes in three different resolutions and spike frame counts, totaling over 9.6k flow fields and 193k+ spike frames. The evaluation dataset consists of 11 scenes, with the first 200 flow fields selected from each scene. To standardize the evaluation data, a center crop is applied to ensure each spike frame and flow field has a width of 1024, with image heights exceeding 768 being cropped to 768. The total number of flow fields and spike frames in the evaluation dataset are 2.2k and 44.22k, respectively.
数据集概述
数据集名称
Real Scenes with Spike and Flow (RSSF)
数据集用途
用于训练和评估基于脉冲的光流。
数据集来源
基于Slow Flow dataset,由高速摄像机捕捉的数据生成。
数据集组成
- 训练集:包含31个场景,分为三种不同分辨率和脉冲帧数的场景。训练集总计有9.6k+流场和193k+脉冲帧。
- 测试集:包含10个场景,每个场景选择前200个流场。测试集总计有2.2k流场和44.22k脉冲帧。
数据集特点
- 训练集和测试集的图像高度超过768的,均被裁剪至768像素;宽度统一为1024像素。
- 流场数量统计仅考虑$dt=20$的情况,$dt=40$和$dt=60$的流场数量与$dt=20$相似。
数据集下载
链接:Baidu NetDisk 密码:2728
数据集部署
- 数据集仅发布脉冲和图像,流场需本地生成。
- 环境要求:Python 3.9,PyTorch 1.10.1+cu113。
- 流场生成步骤:使用提供的
generate_flow.py脚本,根据不同的时间间隔(20, 40, 60)生成流场。
引用信息
@inproceedings{zhao2022learning, title={Learninng optical flow from continuous spike streams}, author={Zhao, Rui and Xiong, Ruiqin and Zhao, Jing and Yu, Zhaofei and Fan, Xiaopeng and Huang, Tiejun}, booktitle={Proceedings of the Annual Conference on Neural Information Processing Systems (NeurIPS)}, year={2022} }




