Image auxiliary annotation method

Abstract

The invention discloses an image auxiliary labeling method, which comprises the following steps of: 1, labeling possible candidate frames on an image by using a pre-trained Yolov5 model, and 2, sending a candidate frame containing a target, a candidate frame region image in a point and point cloud data of the candidate frame region image into a second point cloud classification Point Net + + model, the model automatically identifies a target category according to the image texture and the point cloud data and marks a new rectangular frame, and the image pre-marked in the third step is sent to manual auditing to modify wrong marks, and finally marked data is obtained. In this way, the accuracy of small area labeling during pre-labeling is greatly improved.

Description

technical field [0001] The invention relates to the technical fields of artificial intelligence such as computer vision and deep learning, and in particular to an image-assisted labeling method. Background technique [0002] In recent years, the application of deep learning methods in the field of computer vision has become more and more common. However, deep learning methods are data-driven and need to rely on a large amount of labeled data for model training, which has led to the need for a large amount of data labeling. Due to the large amount of data required, data annotation requires a lot of labor costs and is very time-consuming. If an error is found in the data labeling during the model development process, the data needs to be fed back to the labeler for correction. This process is very long and seriously affects the development efficiency of the model. [0003] The traditional manual labeling method requires a large number of labelers, which is time-consuming and ...

AI Technical Summary

Problems solved by technology

However, in these methods, the training and calling of models, and the inspection of pre-labeled images are often independent and fragmented.

And it is difficult to detect small target objects is the problem of the target detection algorithm. The target detection model used for pre-labeling also exists. The model has a large number of missed detections and false detections for some objects with large scale changes and some small targets, which is not conducive to reducing the difficulty of labeling.

Images