Monocular depth estimation system training method and network based on normalized regression function
A regression function and depth estimation technology, applied in the field of depth estimation, can solve problems such as training loss imbalance and achieve the effect of enhancing system performance
Inactive Publication Date: 2021-04-02
绍兴市北大信息技术科创中心 +2
View PDF2 Cites 6 Cited by
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
[0004] The technical problem to be solved by the present invention is that the training loss of the existing monocular depth estimation system is unbalanced
Method used
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View moreImage
Smart Image Click on the blue labels to locate them in the text.
Smart ImageViewing Examples
Examples
Experimental program
Comparison scheme
Effect test
Embodiment 1
[0039] Such as figure 1 As shown, a training method for monocular depth estimation system based on normalized regression loss function.
[0040] Step 1, preprocessing the training data. Select the required data set for training, such as KITTI unmanned driving data set, Cityscape unmanned driving data set and other public data sets. This embodiment selects the KITTI unmanned driving data set. The image resolution in this data set can be any resolution resolution, such as 1024×960, 1080×600, 960×480, etc. In this embodiment, an image with a resolution of 1024×320 is selected. Randomly read one or more pairs of binocular images from the dataset (I l , I r ), use the stereo matching SGBM algorithm to preprocess the binocular image pair, and get the left eye image disparity map z l , and then use the disparity depth conversion formula to transform the disparity map z l Convert to depth map d′ l , the parallax depth conversion formula is as follows:
[0041]
[0042] Where...
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More PUM
Login to View More Abstract
A monocular depth estimation system training method based on a normalized regression function comprises the following steps: reading a binocular image from a data set, processing a depth map by usinga stereo matching algorithm, selecting a target image from the binocular image of the data set, inputting the target image into a training network, and outputting a predicted depth map, performing regression supervision on the prediction depth map and the proxy depth label by utilizing a regression function, projecting pixel points of the target image into a reference image, sampling and synthesizing new pixels from a projection position in the reference image, generating a reconstructed picture, and calculating an image similarity error between the target image and the reconstructed picture through a luminosity error loss function, thereby obtaining a target image; calculating a smooth error of the prediction depth map through a depth continuity loss function, performing iterative optimization until the luminosity error loss function, the depth continuity loss function and the regression function are converged, and completing training.
Description
technical field [0001] The invention belongs to the field of depth estimation in the field of computer vision, in particular to a training method and network of a monocular depth estimation system based on a normalized regression function. Background technique [0002] Monocular depth estimation is an important research topic in the field of computer vision, and it has a wide range of applications in the fields of robotics, autonomous driving, and augmented reality. In recent years, some systems that use unsupervised methods for monocular depth estimation have emerged. Input binocular image pairs. These methods first use deep neural networks to predict the depth of the target image, and use the predicted depth map and another reference image to reconstruct Generate a new target image. The photometric error between the reconstructed image and the original image is used to optimize the entire deep neural network. In addition, some works use the traditional stereo matching al...
Claims
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More Application Information
Patent Timeline
Login to View More Patent Type & Authority Applications(China)
IPC IPC(8): G06T7/593
CPCG06T7/593G06T2207/10012G06T2207/20081G06T2207/20084G06T2207/20228G06T2207/10028
Inventor 李承远
Owner 绍兴市北大信息技术科创中心