Method and apparatus for obtaining training data

Abstract

The invention provides a method and apparatus for obtaining training data. The method comprises the following steps: receiving original video data; detecting a target object in the original video data by use of a trained first neural network, and automatically marking the detected target object; and based on the automatically marked video data, obtaining the training data used for training a second neural network. According to the method and apparatus for obtaining the training data, provided by the embodiments of the invention, the target object in the original video data is automatically marked by use of the trained neural network, for obtaining the training data used for training a target neural network, such that enormous time spent by manual marking personnel in performing manual marking on the original video data can be saved, the problem of possible omission of the target object due to the manual marking can also be effectively avoided, and high-quality training data is obtained efficiently.

Description

technical field [0001] The present invention relates to the technical field of video processing, and more particularly to a method and device for acquiring training data. Background technique [0002] Neural networks, which play a vital role in the field of computer vision (such as face recognition, object detection, autonomous driving, etc.), have become the dominant technology in this field. Before the neural network can be put into use, it is necessary to use a large number of labeled pictures as training samples to train the neural network. In most cases, the amount and quality of training data has a significant impact on the performance of a neural network. Therefore, large-scale and high-quality training data acquisition is very important. [0003] For neural networks that use video streams as training data, the existing data acquisition method is to pre-record multiple videos, and then send the videos to annotators for frame-by-frame annotation. This method mainly ...

AI Technical Summary

Problems solved by technology

This method mainly has the following problems: (1) in order to obtain enough relevant target objects (such as vehicles, pedestrians, etc.), a large number of videos need to be recorded, and the pre-recorded videos occupy a large amount of storage space; (2) in many scenes, the video Most of the content does not contain relevant target objects, and processing these empty data will take up a lot of time for the annotators; (3) In natural scenes, affected by lighting and other effects, the annotators will miss the label; (4) The annotators are accurate Labeling individual goals takes a lot of effort and time

Images