Pixel-level relative-absolute depth conversion method based on radar point cloud

Abstract

The application discloses a pixel-level relative-absolute depth conversion method based on a radar point cloud. According to the structural relationship between a relative depth map and a radar point cloud map, the application carries out radar point cloud sampling based on pseudo depth, and through the weighted estimation of a small amount of high correlation radar point cloud points to a mapping parameter matrix, the relative-absolute depth conversion is realized. However, due to the pixel-by-pixel estimation and global sampling, the calculation efficiency is greatly reduced, and further, a multi-scale parameter estimation and dynamic neighborhood sampling optimization method based on a DBSCAN algorithm are proposed, so that the efficiency is greatly improved with a very small performance loss. The depth conversion algorithm and the optimization method are combined with each other, the high-performance zero-sample radar point cloud completion effect is realized, and good performances are achieved in structural consistency and depth accuracy. The method corrects the relative depth estimation error, and as much as possible, the good relative depth structural information is retained, so that a high-precision depth completion result map with clear structure is obtained.

Description

Technical Field

[0001] This invention relates to a pixel-level relative-to-absolute depth conversion method based on radar point clouds. The method of this invention converts relative depth into absolute depth, resulting in a dense absolute depth map. The absolute depth map has been greatly improved in terms of structural consistency and depth accuracy. Background Technology

[0002] With the rapid development of the automotive manufacturing industry, the demand for intelligent driving has become increasingly prominent. Depth estimation technology is a core requirement of intelligent driving, and the mainstream depth estimation technology at present is the depth completion method based on radar point cloud signals. Radar point cloud signal-based depth completion methods are divided into supervised radar point cloud signal depth completion algorithms and unsupervised radar point cloud signal depth completion algorithms.

[0003] Supervised radar point cloud signal depth completion algorithms are based on supervised learning of neural networks. They utilize radar signal datasets containing real depth maps to enable the neural network to learn depth completion capabilities. For example, Jinsun Park et al. proposed a LiDAR-based depth completion method in their paper "Non-Local Spatial Propagation Network for Depth Completion," ComputerVision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIII 16. Springer International Publishing, 2020. This method is trained under supervision on the KITTI and NYUv2 datasets, iteratively acquiring dense depth images based on the diffusion of real LiDAR depth points. However, due to the limited coverage of the datasets themselves, supervised radar point cloud signal depth completion algorithms often suffer from low generalization, resulting in poor depth completion performance for scenes not yet present in the dataset.

[0004] Unsupervised radar point cloud signal depth completion algorithms, on the other hand, efficiently estimate the depth corresponding to missing radar depth points by utilizing the depth and spatial relationships of the radar point cloud signal itself and by applying the smoothing prior assumption of the depth map. For example, Y. Zhao et al. proposed an unsupervised depth completion method based on the distribution characteristics of the radar point cloud itself in the paper "A Surface Geometry Model for LiDAR Depth Completion, IEEE Robotics and Automation Letters, 6(3), 4457-4464.". This method effectively achieves radar point cloud depth completion without training samples by pre-filtering the radar point cloud and using depth estimation constraints based on surface normals. However, due to the lack of information from real depth images, the depth estimation accuracy and consistency are relatively poor. Summary of the Invention

[0005] The purpose of this invention is to break down radar point cloud completion into two steps: relative depth estimation and relative-to-absolute depth conversion. Relative depth estimation is not within the scope of this invention; this invention focuses on the implementation of pixel-level relative-to-absolute depth conversion based on radar point clouds. This invention provides a new paradigm for radar point cloud depth completion, enabling pixel-level relative-to-absolute depth conversion. While correcting relative depth estimation errors, it preserves as much relative depth structure information as possible, resulting in a high-precision depth completion image with a clear structure.

[0006] The technical solution adopted in this invention is as follows:

[0007] A pixel-level relative-to-absolute depth conversion method based on radar point clouds is proposed. The method is as follows: based on the relative depth map and the radar point cloud map, the linear mapping parameter matrix is ​​estimated, and the relative-to-absolute depth conversion is achieved by using it. While preserving the structural features of the relative depth, the accurate absolute depth scale is obtained.

[0008] The specific method for estimating the mapping parameter matrix is ​​as follows: calculate the linear mapping parameters corresponding to the pixels in the relative depth map to form a linear mapping parameter matrix; for any pixel in the relative depth map, the method for calculating its linear mapping parameters is as follows: sample the radar point cloud points adjacent to the corresponding pixel in the radar point cloud map using a pseudo-depth-based neighborhood sampling algorithm, and estimate the linear mapping parameters of the pixel based on the relative depth value of the pixel and the absolute depth values ​​of the sampled radar point cloud points adjacent to the pixel; the pseudo-depth-based neighborhood sampling algorithm is as follows: convert the traditional two-dimensional Euclidean distance into a pseudo-three-dimensional distance using relative depth, and perform neighborhood sampling based on the pseudo-three-dimensional distance.

[0009] The specific steps of the above method are as follows:

[0010] First, relative depth is obtained from the visible light image (specifically, existing monocular depth methods or other methods can be used). This relative depth only contains accurate depth structure information but lacks depth scale information.

[0011] Ideally, there should be a globally linear mapping between relative depth and absolute depth, which can be expressed as:

[0012]

[0013] in, Indicates absolute depth. Represents the global linear mapping parameters. and These represent the slope and intercept, respectively. This represents relative depth. Ideally, a relative-to-absolute depth conversion method only requires estimating these two parameters. However, current methods (such as monocular depth estimation) cannot obtain a completely ideal relative depth. To give the relative-to-absolute depth conversion step some degrees of freedom to eliminate the difference between non-ideal relative and absolute depths, the relationship between non-ideal relative and absolute depths is defined as a locally linear mapping relationship, which can be expressed as:

[0014] (1)

[0015] in, and The linear mapping parameter matrix represents the pixel-wise linear mapping, corresponding to the slope and intercept matrices, respectively. This represents the matrix dot product. By estimating the pixel-level linear mapping matrix, depth transformations with non-ideal relative depths can be effectively adapted. However, pixel-level depth transformations lead to a quadratic increase in computational cost. To address this issue, this invention proposes a multi-scale parameter estimation method, adding a grid to the relative depth map, with the distance between the grids being... For each pixel, only the step size is calculated. The linear parameters at the grid points are then upsampled to the original scale using bilinear interpolation, which can be expressed as:

[0016] (2)

[0017] (3)

[0018] in, and Indicates step size is The matrix consists of the linear mapping parameters corresponding to the grid points, where w and h are the width and height of the relative depth map, respectively. This method improves computational efficiency. Multi-scale parameter estimation is based on the observation of continuous parameter estimation. It transforms the full-image mapping parameter estimation into a smaller parameter estimation calculation at the pixel positions corresponding to the downsampled grid, and uses bilinear interpolation algorithm for upsampling, which greatly improves the computational efficiency of the algorithm with less accuracy loss.

[0019] To estimate the linear mapping parameters for corresponding pixels, this invention performs neighborhood sampling on each pixel for which parameters are to be estimated. The neighborhood sampling is limited to a small range, extracting K nearest radar point cloud points, and then selecting the closest point to the given pixel. Using radar point cloud points as depth reference points, the two-dimensional Euclidean distance cannot be completely equivalent to the corresponding depth correlation. Directly sampling from the neighborhood based on distance implies a strong correlation between the similarity between actual depths and their corresponding Euclidean distances. However, in edge regions, some actual depth similarities are not correlated with Euclidean distances, which leads to biases in the final linear mapping matrix estimation and deteriorates the final result. Therefore, this invention further proposes a neighborhood sampling algorithm based on pseudo-depth. By extending the two-dimensional Euclidean distance to a pseudo-three-dimensional distance using the existing relative depth map, neighborhood sampling is performed based on the pseudo-three-dimensional distance. This effectively eliminates misleading radar depth points that are close in Euclidean distance but have low actual depth similarity. The mapping parameter matrix is ​​then weighted and estimated using a small number of highly correlated radar point cloud points. For a single pixel... and radar point cloud The pseudo-3D distance between the two The definition is as follows:

[0020] (4)

[0021] in, This represents the relative depth value at the corresponding location. Based on pseudo-3D distance, the KNN algorithm is used to obtain the nearest distance for each pixel. Using radar points as depth reference points, this process can be represented as:

[0022] (5)

[0023] (6)

[0024] in, Represents pixels The closest pseudo-3D distance A collection of radar point clouds Represents pixels The set of relative depth values, This represents a relative depth map. Furthermore, in order to eliminate... This hyperparameter suffers from insufficient generalization under different conditions, so the DBSCAN algorithm is used to address this issue. Outlier detection is performed on each radar point cloud result. The most concentrated cluster in the radar point cloud set is used as a reference set of high-similarity radar point cloud points, thereby effectively removing interference from low-correlation radar point cloud points. This process can be represented as:

[0025] (7)

[0026] in, This represents a set of reference points for highly similar radar point clouds.

[0027] Based on the obtained high-similarity radar point cloud reference set, the linear mapping parameters corresponding to a pixel can be estimated using weighted least squares regression. In this invention, the weights in the weighted least squares regression are defined with reference to pseudo-3D distance, ensuring that the weight definition conforms to depth correlation, thereby obtaining a depth estimation result that more closely reflects reality. The specific process can be expressed as follows:

[0028] (8)

[0029] (9)

[0030] (10)

[0031] Equation (8) is the weight definition based on pseudo-3D distance, Equation (9) is the definition of the least squares optimization problem, and Equation (10) is the solution to the optimization problem. The weights are defined based on pseudo-3D distance, where This represents the set of highly similar radar point cloud reference points corresponding to the lidar point whose linear mapping parameters are to be estimated. One lidar point, Let the pixel coordinates be the linear mapping parameters to be estimated. This represents the variance control parameter. and They represent estimates respectively. The linear mapping parameter of the position is the first The absolute and relative depth values ​​corresponding to each adjacent lidar point. express Linear mapping parameters of position, and Represents the slope component and intercept component. This represents the weight matrix corresponding to the reference set of highly similar radar point cloud points during the current linear mapping parameter estimation process; and They represent The absolute and relative depth value matrix of the reference set of high-similarity radar point cloud points corresponding to the location of the lidar point.

[0032] Substituting the corresponding values ​​(i.e., the relative depth values ​​of the pixels in the relative depth map and the absolute depth values ​​of the radar point cloud points that are adjacent to the pixel) into equation (10) will yield the final linear mapping parameter matrix, and thus obtain the pixel-level relative-absolute depth conversion.

[0033] The beneficial effects of this invention are as follows:

[0034] The pixel-level relative-to-absolute depth conversion algorithm based on radar point clouds in this invention can be combined with current monocular depth estimation methods that have good performance to obtain the corresponding absolute depth map, thereby improving the depth completion effect and generalization of radar point cloud-based methods. It reduces the difficulty of absolute depth estimation to the simpler single-image relative depth estimation, significantly lowering the problem's complexity and thus greatly improving the method's performance. Based on a multi-scale parameter estimation method, the DBSCAN algorithm is used to dynamically sample neighboring radar point cloud points on a downsampled grid. A pseudo-depth-based neighborhood sampling algorithm assigns pseudo-3D distance weights based on relative depth, performs local linear mapping parameter matrix estimation, and transforms the final result into a full-image linear mapping parameter matrix through bilinear interpolation. Finally, the relative depth to absolute depth conversion is achieved through linear mapping. This invention proposes two efficiency optimization techniques: multi-scale parameter estimation and dynamic neighborhood sampling based on the DBSCAN algorithm, achieving a significant efficiency improvement with minimal performance loss. By combining the depth conversion algorithm with optimization techniques, high-performance zero-sample radar point cloud completion is achieved, particularly in terms of structural consistency and depth accuracy. Attached Figure Description

[0035] Figure 1 Flowchart of a pixel-level relative-absolute depth conversion method based on radar point clouds;

[0036] Figure 2 (a) Visible light image, (b) LiDAR visualization, (c) Relative depth map, (d) Absolute depth conversion result;

[0037] Figure 3 (a) Visible light image, (b) LiDAR visualization, (c) Depth completion result of unsupervised LiDAR proposed by Y. Zhao et al., and (d) Absolute depth conversion result of the present invention. Detailed Implementation

[0038] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0039] like Figure 1 As shown, the pixel-level relative-to-absolute depth conversion method based on radar point clouds of the present invention includes the following steps:

[0040] Based on the relative depth map and radar point cloud map, the linear mapping parameter matrix is ​​estimated, and the relative-to-absolute depth conversion is achieved. While preserving the structural features of the relative depth, the accurate absolute depth scale is obtained.

[0041] The specific method for estimating the mapping parameter matrix is ​​as follows: calculate the linear mapping parameters corresponding to the pixels in the relative depth map to form a linear mapping parameter matrix; for any pixel in the relative depth map, the method for calculating its linear mapping parameters is as follows: sample the radar point cloud points adjacent to the corresponding pixel in the radar point cloud map using a pseudo-depth-based neighborhood sampling algorithm, and estimate the linear mapping parameters of the pixel based on the relative depth value of the pixel and the absolute depth values ​​of the sampled radar point cloud points adjacent to the pixel; the pseudo-depth-based neighborhood sampling algorithm is as follows: convert the traditional two-dimensional Euclidean distance into a pseudo-three-dimensional distance using relative depth, and perform neighborhood sampling based on the pseudo-three-dimensional distance.

[0042] The specific steps of the above method are as follows:

[0043] First, relative depth is obtained from the visible light image (specifically, existing monocular depth methods or other methods can be used). This relative depth only contains accurate depth structure information but lacks depth scale information.

[0044] In a specific embodiment of the present invention, monocular depth estimation is used to obtain... Figure 2 The visible light image shown in (a) The relative depth map is obtained from this. This invention references the method used by L. Yang et al. in the paper "Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data, in Proceedings of the IEEE / CVFConference on Computer Vision and Pattern Recognition, pp. 10371–10381, 2024," and obtains a large-scale monocular depth estimation model by training it on multiple datasets. A relative depth map is obtained based on this model. ,like Figure 2 As shown in (c). This process can be represented as:

[0045]

[0046] in, This represents a relative depth map.

[0047] Ideally, there should be a globally linear mapping between relative depth and absolute depth, which can be expressed as:

[0048]

[0049] in, Indicates absolute depth. Represents the global linear mapping parameters. and These represent the slope and intercept, respectively. This represents relative depth. Ideally, a relative-to-absolute depth conversion method only requires estimating these two parameters. However, current methods (such as monocular depth estimation) cannot obtain a completely ideal relative depth. To give the relative-to-absolute depth conversion step some degrees of freedom to eliminate the difference between non-ideal relative and absolute depths, the relationship between non-ideal relative and absolute depths is defined as a locally linear mapping relationship, which can be expressed as:

[0050] (1)

[0051] in, and The linear mapping parameter matrix represents the pixel-wise linear mapping, corresponding to the slope and intercept matrices, respectively. This represents the matrix dot product. By estimating the pixel-level linear mapping matrix, depth transformations with non-ideal relative depths can be effectively adapted. However, pixel-level depth transformations lead to a quadratic increase in computational cost. To address this issue, this invention proposes a multi-scale parameter estimation method, adding a grid to the relative depth map, with the distance between the grids being... For each pixel, only the step size is calculated. The linear parameters at the grid points are then upsampled to the original scale using bilinear interpolation, which can be expressed as:

[0052] (2)

[0053] (3)

[0054] in, and Indicates step size is The matrix consists of the linear mapping parameters corresponding to the grid points, where w and h are the width and height of the relative depth map, respectively. This method improves computational efficiency. Multi-scale parameter estimation is based on the observation of continuous parameter estimation. It transforms the full-image mapping parameter estimation into a smaller parameter estimation calculation at the pixel positions corresponding to the downsampled grid, and uses bilinear interpolation algorithm for upsampling, which greatly improves the computational efficiency of the algorithm with less accuracy loss.

[0055] To estimate the linear mapping parameters for corresponding pixels, this invention performs neighborhood sampling on each pixel for which parameters are to be estimated. The neighborhood sampling is limited to a small range, extracting K nearest radar point cloud points, and then selecting the closest point to the given pixel. Using radar point cloud points as depth reference points, the two-dimensional Euclidean distance cannot be completely equivalent to the corresponding depth correlation. Directly sampling from the neighborhood based on distance implies a strong correlation between the similarity between actual depths and their corresponding Euclidean distances. However, in edge regions, some actual depth similarities are not correlated with Euclidean distances, which leads to biases in the final linear mapping matrix estimation and deteriorates the final result. Therefore, this invention further proposes a neighborhood sampling algorithm based on pseudo-depth. By extending the two-dimensional Euclidean distance to a pseudo-three-dimensional distance using the existing relative depth map, neighborhood sampling is performed based on the pseudo-three-dimensional distance. This effectively eliminates misleading radar depth points that are close in Euclidean distance but have low actual depth similarity. The mapping parameter matrix is ​​then weighted and estimated using a small number of highly correlated radar point cloud points. For a single pixel... and radar point cloud The pseudo-3D distance between the two The definition is as follows:

[0056] (4)

[0057] in, This represents the relative depth value at the corresponding location. Based on pseudo-3D distance, the KNN algorithm is used to obtain the nearest distance for each pixel. Using radar points as depth reference points, this process can be represented as:

[0058] (5)

[0059] (6)

[0060] in, Represents pixels The closest pseudo-3D distance A collection of radar point clouds Represents pixels The set of relative depth values, This represents a relative depth map. Furthermore, in order to eliminate... This hyperparameter suffers from insufficient generalization under different conditions, so the DBSCAN algorithm is used to address this issue. Outlier detection is performed on each radar point cloud result. The most concentrated cluster in the radar point cloud set is used as a reference set of high-similarity radar point cloud points, thereby effectively removing interference from low-correlation radar point cloud points. This process can be represented as:

[0061] (7)

[0062] in, This represents a reference set of radar point clouds with high similarity.

[0063] Based on the obtained high-similarity radar point cloud reference set, the linear mapping parameters corresponding to a pixel can be estimated using weighted least squares regression. In this invention, the weights in the weighted least squares regression are defined with reference to pseudo-3D distance, ensuring that the weight definition conforms to depth correlation, thereby obtaining a depth estimation result that more closely reflects reality. The specific process can be expressed as follows:

[0064] (8)

[0065] (9)

[0066] (10)

[0067] Equation (8) is the weight definition based on pseudo-3D distance, Equation (9) is the definition of the least squares optimization problem, and Equation (10) is the solution to the optimization problem. The weights are defined based on pseudo-3D distance, where This represents the set of highly similar radar point cloud reference points corresponding to the lidar point whose linear mapping parameters are to be estimated. One lidar point, Let the pixel coordinates be the linear mapping parameters to be estimated. This represents the variance control parameter. and They represent estimates respectively. The linear mapping parameter of the position is the first The absolute and relative depth values ​​corresponding to each adjacent lidar point. express Linear mapping parameters of position, and Represents the slope component and intercept component. This represents the weight matrix corresponding to the reference set of highly similar radar point cloud points during the current linear mapping parameter estimation process; and They represent The absolute and relative depth value matrix of the reference set of high-similarity radar point cloud points corresponding to the location of the lidar point.

[0068] Substituting the corresponding values ​​(i.e., the relative depth values ​​of the pixels in the relative depth map and the absolute depth values ​​of the radar point cloud points that are adjacent to the pixel) into equation (10) will yield the final linear mapping parameter matrix, and thus obtain the pixel-level relative-absolute depth conversion.

[0069] In one specific embodiment of the present invention, radar point clouds are acquired by a radar instrument and projected onto a two-dimensional plane to form a radar point cloud map. Radar point cloud sets are formed on a downsampling grid with a step size of 4, based on pseudo-3D distance sampling of neighborhood radar point cloud points. and the corresponding set of relative depth values. Based on the weights derived from pseudo-3D distance, the relative depth values ​​of pixels in the relative depth map and the absolute depth values ​​of radar point cloud points sampled from neighboring pixels are substituted into the formula corresponding to the solution of the weighted linear estimation to obtain the pixel-by-pixel linear mapping parameters. Finally, the full-scale linear mapping parameter matrix is ​​obtained based on bilinear interpolation. This achieves a relative-to-absolute depth conversion result, such as... Figure 2 As shown in (d). And Figure 2 Compared with (a), the method of the present invention has achieved good results in terms of depth accuracy and consistency.

[0070] The following compares the absolute depth map obtained using the method of this invention with the absolute depth map results obtained based on other radar point cloud depth completion models from a visual perspective. Figure 2 As can be seen, (a) is a visible light image, (b) is a LiDAR visualization image, (c) is a relative depth map obtained from the monocular depth estimation model, and (d) is the absolute depth map after transformation by this invention. It can be seen that this invention can effectively achieve accurate relative-to-absolute depth transformation based on radar point clouds, restoring relatively accurate scale information while ensuring structural consistency. Figure 3 It can be further seen that (a) is a visible light image, (b) is a lidar image, (c) is the radar point cloud depth completion result proposed by Y. Zhao et al., and (d) is the absolute depth map result transformed by the method of this invention. It can be seen that the pixel-level relative-absolute depth conversion method based on radar point clouds of this invention significantly outperforms other methods in terms of depth accuracy and consistency. Compared to (c), the result generated by the method of this invention shows superior performance, especially in terms of object depth consistency and depth accuracy.

[0071] The above are merely specific embodiments of the present invention and should not be construed as limiting the scope of the present invention. Equivalent variations made by those skilled in the art based on this invention, as well as changes well-known to those skilled in the art, should still fall within the scope of the present invention.

Claims

1. A pixel-level relative-to-absolute depth conversion method based on radar point clouds, characterized in that: Based on the relative depth map and radar point cloud map, a linear mapping parameter matrix is ​​estimated, and this is used to achieve relative-to-absolute depth conversion; among which, The specific method for estimating the linear mapping parameter matrix is ​​as follows: Calculate the linear mapping parameters corresponding to the pixels in the relative depth map to form a linear mapping parameter matrix; for any pixel in the relative depth map, the calculation method for its linear mapping parameters is as follows: sample the radar point cloud points adjacent to the pixel's corresponding point cloud point in the radar point cloud map using a pseudo-depth-based neighborhood sampling algorithm; estimate the pixel's linear mapping parameters based on the pixel's relative depth value and the sampled absolute depth values ​​of the adjacent radar point cloud points; the pseudo-depth-based neighborhood sampling algorithm is as follows: transform the traditional two-dimensional Euclidean distance into a pseudo-three-dimensional distance using relative depth, and perform neighborhood sampling based on the pseudo-three-dimensional distance to obtain a high-similarity radar point cloud point reference set; the specific method for transforming the traditional two-dimensional Euclidean distance into a pseudo-three-dimensional distance in the pseudo-depth-based neighborhood sampling algorithm is as follows: For pixels and radar point cloud The pseudo-3D distance between the two The definition is as follows: (1) in, This represents the relative depth value at the corresponding location; The linear mapping parameters of the pixel are estimated based on the relative depth value of the pixel and the absolute depth values ​​of the radar point cloud points adjacent to the pixel obtained by sampling. The specific method is as follows: Based on the obtained high-similarity radar point cloud reference set, the linear mapping parameters corresponding to the pixels are estimated by weighted least squares regression. The weights in the weighted least squares regression are defined by referring to the pseudo-3D distance, so that the weight definition conforms to depth correlation. The estimation process of the linear mapping parameters is expressed as follows: (3) (4) (5) in, For pseudo 3D distance Defined weights, This represents the set of highly similar radar point cloud reference points corresponding to the lidar point whose linear mapping parameters are to be estimated. One lidar point, The pixel coordinates are the linear mapping parameters to be estimated. Let represent the variance control parameter, and Equation (4) defines the least squares optimization problem. and They represent estimates respectively. The linear mapping parameter of the position is the first The absolute and relative depth values ​​corresponding to each adjacent lidar point. express Linear mapping parameters of position, and These represent the slope component and the intercept component, respectively. express The weight matrix of the reference set of highly similar radar point cloud points corresponding to the location of the lidar point; and They represent The absolute depth matrix and relative depth matrix of the reference set of highly similar radar point cloud points corresponding to the location's lidar points.

2. The pixel-level relative-absolute depth conversion method based on radar point clouds according to claim 1, characterized in that: The specific process of neighborhood sampling based on pseudo-3D distance is as follows: For the sampling... Each radar point cloud point is analyzed using the DBSCAN algorithm. The most concentrated cluster in the set of radar point cloud points is taken as the reference set of highly similar radar point cloud points, specifically represented as: (2) in, This represents a reference set of radar point clouds with high similarity.

3. The pixel-level relative-to-absolute depth conversion method based on radar point clouds according to claim 1, characterized in that: In the process of calculating the linear mapping parameters corresponding to pixels in the relative depth map, a multi-scale parameter estimation method is used, and a grid is added to the relative depth map, with the distance between the grids being... Pixels, calculation step size is The linear mapping parameters on the grid points are then upsampled to the original scale using bilinear interpolation.

Images