Depth estimation method and model, and data processing equipment

Abstract

The invention provides a depth estimation method and model, and data processing equipment. The data processing equipment is configured with a convolutional neural network model of a U-Net structure. The method comprises the following steps: obtaining a to-be-processed image, and inputting the to-be-processed image into an encoder; processing the input data in sequence through multiple levels of down-sampling convolution layers in the encoder, and outputting feature map data obtained through processing to a decoder; processing the input feature map data in sequence through multiple levels of up-sampling convolution layers in the decoder, and adding local depth data into the feature map data output by at least one up-sampling convolution layer through a plane fitting layer; obtaining a depth estimation result output by the decoder. Therefore, by adding geometric constraints brought by a fitting plane in features, results obtained by estimation can have a more stable numerical range, so that a depth estimation result can be more accurate.

Description

technical field [0001] The present application relates to the field of image data processing, in particular, to a depth estimation method, model and data processing equipment. Background technique [0002] Image depth estimation is a technique for estimating the physical depth value in an image based on image information on a plane. Depth estimation can be used in scenarios such as 3D modeling, depth perception, and scene understanding. For example, on a mobile terminal, depth estimation can be used to estimate the degree relationship of portraits, objects or backgrounds during shooting, so that different levels of blur can be used according to different depths, so as to achieve a more realistic and natural blur gradient effect. [0003] At present, the depth estimation is mainly realized by the machine learning model trained according to a large number of samples. Among them, the machine learning model trained by unsupervised learning has poor generalization ability on dif...

AI Technical Summary

Problems solved by technology

[0003] At present, the depth estimation is mainly realized by the machine learning model trained according to a large number of samples. Among them, the machine learning model trained by unsupervised learning has poor generalization ability on different depth data sets and is not universal; The main data of the machine learning model trained by supervised learning is used to directly fit the neural network, which lacks the constraints and correspondence of certain geometric relationships, resulting in inaccurate depth estimation results.

Images