Interactive intelligent 2D semantic segmentation system and method, storage medium and device

Abstract

The invention provides an interactive intelligent 2D semantic segmentation system. The system comprises an image input interface, an annotation front end and an annotation rear end; the image input interface is connected with a target image library, the image input interface is connected with the annotation rear end, the annotation rear end receives a target image through the image input interface, the annotation rear end is connected with the annotation front end, and the invention is characterized in that the annotation rear end receives an annotation range and an annotation type of the annotation front end; the annotation rear end comprises a template storage module, a cutting module, a foreground extraction module and a contour extraction module; the cutting module and the template storage module are connected with the foreground extraction module; the cutting module cuts the target image; the template storage module is used for producing a template; the foreground extraction module is connected with the contour extraction module; the foreground extraction module generates a foreground extraction template; the contour extraction module extracts an image contour. The beneficialeffects are that the interactive intelligent 2D semantic segmentation system is provided, the ground-truth can be conveniently obtained, and the robustness and the accuracy are relatively high.

Description

technical field [0001] The invention relates to data labeling, in particular to an interactive intelligent 2D semantic segmentation system, method, storage medium and device. Background technique [0002] The traditional 2D semantic segmentation labeling method uses completely manual labeling, and the labeler uses polygons to completely surround the picture pixels to which the target object belongs, so as to achieve the effect of 2D semantic segmentation. The labeling accuracy of the 2D semantic segmentation labeling task depends on the degree of fit between the polygon and the outline of the target object. In most cases, the higher the degree of conformity with the contour of the target object is required, the more polygon vertices are required. In order to achieve higher labeling accuracy, the labeler needs to draw polygons with more vertices to fit the outline of the target object. The more complex the outline of the target object, the more difficult the labeling will be...

AI Technical Summary

Problems solved by technology

In order to achieve higher labeling accuracy, the labeler needs to draw polygons with more vertices to fit the outline of the target object. The more complex the outline of the target object, the more difficult the labeling will be, and the lower the labeling efficiency will be.

It can be seen that the traditional completely artificial 2D semantic segmentation method has low labeling efficiency and difficult to improve accuracy.

[0003] Although CN108734113A proposes a method for automatic vehicle labeling, on the one hand, as a semi-supervised object segmentation, the ground-truth of the first frame of the object is required, which is troublesome to obtain in practice; on the other hand, it lacks interactive functions, resulting in a lack of robustness sex and accuracy

Images