使用视频帧嵌入来训练视频数据生成神经网络
By using video data embedding neural networks to process the embedding of video frames, calculate similarity and update parameters, the problem of evaluating spatiotemporal relationships in neural network training is solved, and the generation of realistic video data is achieved efficiently.
CN116097278BActive Publication Date: 2026-07-17GDM HOLDING LLC
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GDM HOLDING LLC
- Filing Date
- 2021-09-08
- Publication Date
- 2026-07-17
AI Technical Summary
Technical Problem
Training neural networks to generate high-quality video data is difficult, especially since the objective function for evaluating spatiotemporal relationships is difficult to formulate and is not differentiable, resulting in low training efficiency.
Method used
The video data embedding neural network processes the embeddings of training and target video frames, determines the gradient by calculating similarity such as Frechet distance, and updates the parameters of the video data generation network.
Benefits of technology
It achieves efficient training when generating video data, and can generate realistic, continuous or high-resolution video data, and is suitable for video generation that is not time-aligned or does not depict corresponding content.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN116097278B_ABST
Abstract
用于训练具有多个视频生成网络参数的视频数据生成神经网络的方法、系统和装置，包括在计算机存储介质上编码的计算机程序。在一个方面中，一种方法包括：根据视频数据生成网络参数的当前值使用视频数据生成神经网络来生成一个或多个训练视频帧序列；获得一个或多个目标视频帧序列；以及使用从训练视频帧和目标视频帧的相应嵌入之间的相似度导出的训练信号来训练视频数据生成神经网络。嵌入是由视频数据嵌入神经网络生成的。
Need to check novelty before this filing date? Find Prior Art