A semantic annotation method for indoor scenes based on rgb‑d data

Abstract

The invention relates to an indoor scene semantic labeling method based on RGB-D data, constructs a semantic labeling framework based on RGB-D data from coarse to fine global recursive feedback, and divides the entire semantic labeling framework into coarse-grained Region-level semantic label inference and fine-grained pixel-level semantic label refinement are two parts; unlike the traditional single region-level or pixel-level semantic labeling framework, this framework re-establishes coarse-grained region-level semantic labeling and fine-grained pixel-level semantic labeling By introducing a reasonable global recursive feedback mechanism, the coarse-grained region-level semantic annotation results and the fine-grained pixel-level semantic annotation results are alternately iteratively updated and optimized. In this way, the multi-modal information of different regional levels in the scene image is well integrated, and to a certain extent, it solves the problem that it is difficult to properly select the annotation primitives that are common in traditional indoor scene semantic annotation schemes.

Description

technical field [0001] The invention relates to an image semantic labeling method, in particular to an indoor scene semantic labeling method based on RGB-D data, and belongs to the technical field of semantic label classification of computer vision. Background technique [0002] Image semantic annotation is the core unit of scene understanding work in computer vision, and its basic goal is to densely provide a predefined semantic category label for each pixel in a given query image. Considering the ambiguity, complexity and abstraction of image semantics, the image semantic models generally established are hierarchical. Among them, "target semantics" is at the middle level of the semantic hierarchy, and plays a linking role in many high-level semantic reasoning processes. According to the quantitative level of annotation primitives in image semantic annotation, most current image semantic annotation schemes can be roughly divided into two categories, including: pixel-level ...

AI Technical Summary

Problems solved by technology

However, most current semantic annotation works only use depth information to construct region-level features, but ignore its role in context inference, and the depth information used is relatively simple.

[0007] To sum up, the existing semantic annotation schemes for indoor scenes generally have the problem that it is difficult to select the quantization level of the annotation primitives, and the role of geometric depth information in the context reasoning process has not received enough attention.

Images