V2V数据集
收藏资源简介:
V2V数据集由北京大学的研究团队创建,旨在解决基于事件的视觉模型训练数据稀缺的问题。该数据集包含10,000个视频,总时长为52小时,比现有的基于事件的相机数据集大一个数量级。数据集通过将传统视频帧直接转换为基于事件的体素网格表示,从而消除了对存储密集型事件流生成的需求。V2V方法使得存储需求降低了150倍,并支持在训练过程中的参数随机化,增强了模型的鲁棒性。该数据集可用于事件视频重建和光流估计等任务,为开发鲁棒的基于事件的算法提供了新的可能性。
The V2V dataset was developed by a research team from Peking University to address the scarcity of training data for event-based vision models. It contains 10,000 videos with a total duration of 52 hours, which is one order of magnitude larger than existing event-based camera datasets. By directly converting conventional video frames into event-based voxel grid representations, the dataset eliminates the need for storage-intensive event stream generation. The V2V approach reduces storage requirements by 150 times, supports parameter randomization during training, and thereby enhances model robustness. This dataset can be applied to tasks such as event video reconstruction and optical flow estimation, offering new opportunities for developing robust event-based algorithms.

- 1V2V: Scaling Event-Based Vision through Efficient Video-to-Voxel Simulation北京大学 · 2025年



