Weak supervision method for estimating depth from truncation phase
Through a weak supervision method for estimating depth from truncated phase, the consistency loss supervision technology is used to solve the problem of poor generalization of the depth estimation model in structured light three-dimensional reconstruction, and stable depth recovery under different materials and lighting conditions is achieved. Efficient labeling improves the accuracy and reliability of three-dimensional reconstruction.
Patent Information
- Application Number
- CN202411156053.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-22
- Publication Date
- 2024-11-26
AI Technical Summary
In the existing three-dimensional reconstruction technology of structured light, the neural network depth estimation model is difficult to generalize under different object materials and lighting conditions, resulting in a decrease in depth estimation effect and high cost of labeling training data.
A weak supervision method for estimating depth from truncated phase is adopted, through data preprocessing and consistency loss supervision, depth maps close to truth value are generated, labor costs of training data annotation, and depth is stable under different materials and lighting conditions. .
It improves the accuracy and reliability of structured light three-dimensional reconstruction, reduces the cost of training data annotation, and ensures depth estimation performance in complex scenarios.
Smart Images

Figure CN119027472A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of visual processing technology, and in particular to a weakly supervised method for estimating depth from a truncated phase. Background Art
[0002] In structured light 3D imaging, the method of projecting stripe structured light is often used to obtain the 3D surface information of the object surface. Figure 1 As shown in the figure, the principle is that a projector projects a set of sinusoidally coded structured light stripes onto the surface of the object to be measured. Due to the undulations on the surface of the object, these regular stripes will be distorted and deformed when observed from a specific angle. The camera is used to capture these deformed stripe patterns, and the camera can capture the encoded depth information. The truncated phase can then be decoded through a phase shift algorithm, and then the truncated phase can be restored to obtain the depth information.
[0003] However, in practical applications, fast and high-precision reconstruction is still a challenge faced by structured light 3D reconstruction technology. With the popularization of neural network technology, it has gradually been applied in structured light 3D reconstruction, providing new possibilities for achieving high-speed and high-precision structured light 3D reconstruction. Neural network models can be used to learn complex features and patterns from large amounts of data. Through training, images captured by structured light can be analyzed more accurately, and robustness can be maintained in difficult situations such as lighting changes, occlusions, or strongly reflective surfaces.
[0004] In the current supervised neural network method, in the FPP (FringeProjection Profilometry) system with fixed camera and projector parameters, the depth estimation network trained from a certain type of data performs poorly when processing other data. The fundamental reason is that when scanning different objects, changes in the object material will cause changes in the luminosity distribution of the fringe image, making it difficult for the depth estimation network trained on a single data set to generalize. At the same time, when using supervised methods to train neural networks to achieve structured light 3D reconstruction, the amount of training data that needs to be collected is relatively large, and the training data labels require a lot of manual processing. Summary of the invention
[0005] The invention provided by the present invention aims to provide a weakly supervised method for estimating depth from a truncated phase. Through the weakly supervised method for estimating depth from a truncated phase of the present invention, the weakly supervised method for estimating depth from a truncated phase mainly serves to reduce the labor cost of training data annotation, and solve the problem that the luminance distribution of stripes changes due to different object materials, thereby causing the neural network to be unable to generalize well. In practical applications, the FPP system may encounter different lighting conditions, which will affect the quality and visibility of the stripes. Traditional supervised and weakly supervised depth estimation methods are often difficult to generalize under these changing conditions, that is, the performance of a model trained under one condition may degrade under another condition. This method can stably recover the depth from the truncated phase of objects of different materials, thereby ensuring accuracy and reliability in complex scenes.
[0006] In order to achieve the above effect, the present invention provides the following technical solution: a weakly supervised method for estimating depth from a truncated phase, comprising the following steps: S1. Data preprocessing: recovering the truncated phase image from the fringe image.
[0007] S2. Supervising depth map generation using consistency loss with truncated phase.
[0008] S3. Supervise depth map generation using a one-cycle continuous phase consistency loss.
[0009] S4. Generates a depth map close to the ground truth after supervision using two losses.
[0010] Further, according to the operation steps in S1, S101, first collect 64 cycles of three-step phase-shift fringe patterns from the FPP system , by adjusting the order of the three-step phase-shifted fringe images.
[0011] S102. Calculate three different truncated phase images according to the formula .
[0012] Further, according to the operation steps in S102, the formula In the formula, , , is a 64-cycle three-step phase-shift fringe pattern collected from the FPP system. is the calculated truncated phase map.
[0013] Further, according to the operation steps in S2, S201. The 64-cycle truncated phase under the projector viewing angle is fixed, and the 64-cycle sinusoidal fringe encoding is obtained using the formula.
[0014] S202: Through the preset fringe coding, the 64-cycle truncated phase image under the projector viewing angle can be calculated. .
[0015] S203, input the obtained truncated phase image into the depth estimation network, and then an initial estimated depth map can be obtained. The resolution of the depth map is the same as the input truncated phase image. When the posture relationship between the camera and the projector and the depth D corresponding to each pixel coordinate i, j in the depth map are known, the depth map can be converted into a point cloud in three-dimensional space and reprojected onto the projector plane, so that the pixel coordinate value of each point in the depth map can be calculated. The corresponding pixel coordinate value under the projector's viewing angle .
[0016] S204, reprojecting the 64-cycle truncated phase image from the projector's perspective to the camera's perspective. Since the u and v pixel coordinate values obtained by reprojection are generally floating point numbers, the truncated phase image from the projector's perspective The pixel coordinates are integers, so bilinear interpolation is needed to obtain the corresponding pixel coordinate values. Finally, the 64-cycle truncated phase image under the camera's perspective can be obtained. .
[0017] S205, in getting Then, compare it with the input A supervisory signal is established to construct the truncated phase consistency loss.
[0018] Further, according to the operation steps in S201, the formula is In the formula, , where h is the height of the image, represents the number of pixels in a fringe period, is the sequence number of the three-step phase shift, represents the pixel coordinates on the projector plane, Refers to the projector.
[0019] Further, according to the operation steps in S204, the The calculation formula is In the formula, Represents the pixel coordinate value from the camera's perspective. Indicates the pixel coordinate value under the projector's viewing angle, with the superscript Refers to the projector, Indicates the coordinates under the projector's viewing angle The cutoff phase value of Indicates the coordinates from the camera's perspective The cutoff phase value.
[0020] Further, according to the operation steps in S205, the .
[0021] In the formula, represents the L1 norm loss of two truncated phase images, represents three interpolated truncated phase images from the camera perspective, and Three truncated phases representing the input Figure 1 One to one correspondence.
[0022] Representatives in Point to the structural similarity of the two corresponding truncated phase images, , represents the proportionality coefficient of different losses, Represents the effective point, Represents the set of valid points.
[0023] Further, according to the operation steps in S205, the selection of the effective point can be obtained by the formula: In the formula, the set represents the effective point obtained by setting the modulation threshold value, represents a point in the set, represents the serial number of the truncated phase image, Indicates On the picture The truncated phase value of the point, express The flag to indicate whether the point is selected. 1 means it is selected, and 0 means it is not selected.
[0024] Further, according to the operation steps in S3, S301, first collect a period of fringe pattern from the FPP system .
[0025] S302, using a one-cycle three-step phase-shift fringe pattern, a one-cycle truncated phase pattern can be obtained through a formula.
[0026] S303. After obtaining a truncated phase diagram of one cycle, it can be converted into an absolute phase through a formula.
[0027] S304, using a formula to calculate a one-cycle absolute phase diagram under the camera's viewing angle , the absolute phase of one cycle on the projector plane is known and is expressed as , the positional relationship between the projector and the camera is known. Through the corresponding relationship of the formula, we can get a continuous phase image of a cycle reprojected from the projector's perspective to the camera's perspective. , thus constructing a cycle phase consistency loss.
[0028] S301. Finally, two more losses are added to constrain the obtained depth map comparison.
[0029] Further, according to the operation steps in S302, the formula In the formula, represents a fringe pattern of one cycle, represents the pixel coordinate value on the fringe pattern, Represents the serial number of the fringe pattern.
[0030] According to the operation steps in S303, the formula In the formula, is the pixel coordinate The truncated phase value at is the pixel coordinate calculated according to the formula The absolute phase value at .
[0031] According to the operation steps in S304, the formula In the formula, the set represents the effective point, which is exactly the same as the effective point mentioned above. Indicate point The absolute phase value of one cycle reprojected to the camera plane, Indicates the absolute phase value of the effective point in the camera plane over one cycle.
[0032] The present invention provides a weakly supervised method for estimating depth from a truncated phase, which has the following beneficial effects: the weakly supervised method for estimating depth from a truncated phase mainly reduces the labor cost of training data annotation and solves the problem that the different materials of objects cause the stripe luminosity distribution to change, resulting in the inability of the neural network to generalize well. In practical applications, the FPP system may encounter different lighting conditions, which will affect the quality and visibility of the stripes. Traditional supervised and weakly supervised depth estimation methods are often difficult to generalize under these changing conditions, that is, the performance of a model trained under one condition may degrade under another condition. This method can stably recover the depth from the truncated phase of objects of different materials, thereby ensuring accuracy and reliability in complex scenes. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 A schematic diagram of the principle of structured light three-dimensional imaging of a weakly supervised method for estimating depth from a truncated phase according to the present invention; Figure 2 A schematic diagram of the operation flow of a weakly supervised method for estimating depth from a truncated phase according to the present invention; Figure 3 A schematic diagram of a three-step phase-shifted fringe image and a truncated phase in a weakly supervised method for estimating depth from a truncated phase according to the present invention; Figure 4 A schematic diagram of a three-step phase shift one-cycle fringe image according to a weakly supervised method for estimating depth from truncated phase of the present invention; Figure 5 A schematic diagram of the comparison between depth maps and true values generated under different losses in a weakly supervised method for estimating depth from a truncated phase according to the present invention; Figure 6 A schematic diagram of a depth map reconstruction result of a weakly supervised method for estimating depth from a truncated phase according to the present invention; Figure 7 The figure is a flow chart of a weakly supervised method for estimating depth from a truncated phase according to the present invention. DETAILED DESCRIPTION
[0034] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.
[0035] The present invention provides a technical solution: please refer to Figure 1-7 , a weakly supervised method for estimating depth from truncated phases, comprising the following steps: S1. Data preprocessing: recovering the truncated phase image from the fringe image.
[0036] S2. Supervising depth map generation using consistency loss with truncated phase.
[0037] S3. Supervise depth map generation using a one-cycle continuous phase consistency loss.
[0038] S4. Generates a depth map close to the ground truth after supervision using two losses.
[0039] Specifically, according to the operation steps in S1, S101, first collect 64 cycles of three-step phase-shift fringe patterns from the FPP system , by adjusting the order of the three-step phase-shifted fringe images.
[0040] S102. Calculate three different truncated phase images according to the formula .
[0041] Specifically, according to the operation steps in S102, the formula In the formula, , , is a 64-cycle three-step phase-shift fringe pattern collected from the FPP system. is the calculated truncated phase map.
[0042] Specifically, according to the operation steps in S2, S201. The 64-cycle truncated phase under the projector viewing angle is fixed, and the 64-cycle sinusoidal fringe encoding is obtained using the formula.
[0043] S202: Through the preset fringe coding, the 64-cycle truncated phase image under the projector viewing angle can be calculated. .
[0044] S203, input the obtained truncated phase image into the depth estimation network, and then an initial estimated depth map can be obtained. The resolution of the depth map is the same as the input truncated phase image. When the posture relationship between the camera and the projector and the depth D corresponding to each pixel coordinate i, j in the depth map are known, the depth map can be converted into a point cloud in three-dimensional space and reprojected onto the projector plane, so that the pixel coordinate value of each point in the depth map can be calculated. The corresponding pixel coordinate value under the projector's viewing angle .
[0045] S204, reprojecting the 64-cycle truncated phase image from the projector's perspective to the camera's perspective. Since the u and v pixel coordinate values obtained by reprojection are generally floating point numbers, the truncated phase image from the projector's perspective The pixel coordinates are integers, so bilinear interpolation is needed to obtain the corresponding pixel coordinate values. Finally, the 64-cycle truncated phase image under the camera's perspective can be obtained. .
[0046] S205, in getting Then, compare it with the input A supervisory signal is established to construct the truncated phase consistency loss.
[0047] Specifically, according to the operation steps in S201, the formula is In the formula, , where h is the height of the image, represents the number of pixels in one fringe period, is the sequence number of the three-step phase shift, represents the pixel coordinates on the projector plane, Refers to the projector.
[0048] Specifically, according to the operation steps in S204, the The calculation formula is In the formula, Represents the pixel coordinate value from the camera's perspective. Indicates the pixel coordinate value under the projector's viewing angle, with the superscript Refers to the projector, Indicates the coordinates under the projector's viewing angle The cutoff phase value of Indicates the coordinates from the camera's perspective The cutoff phase value.
[0049] Specifically, according to the operation steps in S205, the .
[0050] In the formula, represents the L1 norm loss of two truncated phase images, represents three interpolated truncated phase images from the camera perspective, and Three truncated phases representing the input Figure 1 One to one correspondence.
[0051] Representatives in Point to the structural similarity of the two corresponding truncated phase images, , represents the proportionality coefficient of different losses, Represents the effective point, Represents the set of valid points.
[0052] Specifically, according to the operation steps in S205, the selection of the effective point can be obtained by the formula: In the formula, the set represents the effective point obtained by setting the modulation threshold. represents a point in the set, represents the serial number of the truncated phase image, Indicates On the picture The truncated phase value of the point, express The flag to indicate whether the point is selected. 1 means it is selected, and 0 means it is not selected.
[0053] Specifically, according to the operation steps in S3, S301, first collect a period of fringe pattern from the FPP system .
[0054] S302, using a one-cycle three-step phase-shift fringe pattern, a one-cycle truncated phase pattern can be obtained through a formula.
[0055] S303. After obtaining a truncated phase diagram of one cycle, it can be converted into an absolute phase through a formula.
[0056] S304, using a formula to calculate a one-cycle absolute phase diagram under the camera's viewing angle , the absolute phase of one cycle on the projector plane is known and is expressed as , the positional relationship between the projector and the camera is known. Through the corresponding relationship of the formula, we can get a continuous phase image of a cycle reprojected from the projector's perspective to the camera's perspective. , thus constructing a cycle phase consistency loss.
[0057] S301. Finally, two more losses are added to constrain the obtained depth map comparison.
[0058] Specifically, according to the operation steps in S302, the formula In the formula, represents a fringe pattern of one cycle, represents the pixel coordinate value on the fringe pattern, Represents the serial number of the fringe pattern.
[0059] According to the operation steps in S303, the formula In the formula, is the pixel coordinate The truncated phase value at is the pixel coordinate calculated according to the formula The absolute phase value at .
[0060] According to the operation steps in S304, the formula In the formula, the set represents the effective point, which is exactly the same as the effective point mentioned above. Indicate point The absolute phase value of one cycle reprojected to the camera plane, Indicates the absolute phase value of the effective point in the camera plane over one cycle.
[0061] The images generated by the embodiment are tested and analyzed, and compared with the prior art, and the following data are obtained: accuracy reliability Convenience Example Higher Higher Higher Prior art generally generally generally According to the data in the above table, it can be concluded that when the embodiment is implemented, a weakly supervised method for estimating depth from a truncated phase of the present invention is used. The weakly supervised method for estimating depth from a truncated phase mainly reduces the labor cost of training data annotation and solves the problem that the luminosity distribution of stripes changes due to different object materials, resulting in the inability of the neural network to generalize well. In practical applications, the FPP system may encounter different lighting conditions, which will affect the quality and visibility of the stripes. Traditional supervised and weakly supervised depth estimation methods are often difficult to generalize under these changing conditions, that is, the performance of a model trained under one condition may degrade under another condition. This method can stably recover the depth from the truncated phase of objects of different materials, thereby ensuring accuracy and reliability in complex scenes.
[0062] The present invention provides a weakly supervised method for estimating depth from truncated phase, comprising the following steps: S1, data preprocessing, recovering the truncated phase image from the fringe image, firstly collecting a 64-cycle three-step phase shift fringe image from the FPP system , , By adjusting the order of the three-step phase shift fringe images, three different truncated phase images can be calculated according to formulas 1-1 1-2 1-3 ; (1-1) (1-2) (1-3) In the formula, , , is a 64-cycle three-step phase-shift fringe pattern collected from the FPP system. is the calculated truncated phase map.
[0063] The final result is as follows Figure 3 The first row shows the 64-cycle three-step phase-shift fringe pattern collected. , , The second line is the three truncated phase diagrams calculated by the formula , the three truncated phase images maintain a relationship similar to a three-step phase shift.
[0064] S2, supervised depth map generation using consistency loss of truncated phase; The 64-cycle truncated phase under the projector's viewing angle is fixed, and the 64-cycle sinusoidal fringe encoding is obtained using formula 2-1: (2-1) In the formula, , where h is the height of the image, represents the number of pixels in a fringe period, is the sequence number of the three-step phase shift. represents the pixel coordinates on the projector plane, Refers to the projector.
[0065] Through the preset fringe coding, the 64-cycle truncated phase image under the projector's viewing angle can be calculated. .
[0066] The obtained truncated phase image is input into the depth estimation network to obtain an initial estimated depth map, and the resolution of the depth map is the same as the input truncated phase image. When the pose relationship between the camera and the projector and the depth D corresponding to each pixel coordinate i and j in the depth map are known, the depth map can be converted into a point cloud in three-dimensional space and reprojected onto the projector plane, so that the pixel coordinate value of each point in the depth map can be calculated. The corresponding pixel coordinate value under the projector's viewing angle .
[0067] After knowing this correspondence, the 64-cycle truncated phase image under the projector's perspective can be reprojected to the camera's perspective. Since the u and v pixel coordinate values obtained by reprojection are generally floating point numbers, the truncated phase image under the projector's perspective The pixel coordinates are integers, so bilinear interpolation is needed to obtain the corresponding pixel coordinate values. Finally, the 64-cycle truncated phase image under the camera's perspective can be obtained. , as shown in Formula 2-2: (2-2) In the formula, Represents the pixel coordinate value from the camera's perspective. Indicates the pixel coordinate value under the projector's viewing angle, with the superscript Refers to the projector, Indicates the coordinates under the projector's viewing angle The cutoff phase value of Indicates the coordinates from the camera's perspective The cutoff phase value.
[0068] In getting Then, compare it with the input (Formulas 1-1, 1-2, 1-3) establish a supervisory signal to construct the truncated phase consistency loss as shown in Formula 2-3: (2-3) In the formula, represents the L1 norm loss of two truncated phase images, represents three interpolated truncated phase images from the camera perspective, and Three truncated phases representing the input Figure 1 One corresponding, Representatives in Point to the structural similarity of the two corresponding truncated phase images, , represents the proportionality coefficient of different losses, Represents the effective point, Represents the set of valid points. The selection of valid points can be obtained by formulas 2-4 and 2-5: (2-4) (2-5) In the formula, the set Indicates the effective point obtained by setting the modulation threshold value. represents a point in the set, represents the serial number of the truncated phase image, Indicates On the picture The cutoff phase value of the point. express The flag to indicate whether the point is selected. 1 means it is selected, and 0 means it is not selected.
[0069] S3, supervise the depth map generation using a one-cycle continuous phase consistency loss; Relying solely on high-frequency truncated phase consistency loss is often not effective in predicting depth, and other constraints need to be introduced. This section introduces the second loss, one-cycle continuous phase consistency loss.
[0070] First, collect a period of fringe patterns from the FPP system ,like Figure 4 As shown: Using a three-step phase shift fringe pattern of one cycle, we can obtain a truncated phase pattern of one cycle through formula 3-1: (3-1) In the formula, represents a fringe pattern of one cycle, represents the pixel coordinate value on the fringe pattern, Represents the serial number of the fringe pattern.
[0071] After obtaining the truncated phase diagram of one cycle, it can be converted into absolute phase using formula 3-2: (3-2) In the formula, is the pixel coordinate The truncated phase value at It is the pixel coordinate calculated according to formula 3-1 The absolute phase value at Formulas 3-1 and 3-2 can be used to calculate the absolute phase diagram of one cycle under the camera's perspective: , the absolute phase of one cycle on the projector plane is known and is expressed as The positional relationship between the projector and the camera is known. Through the corresponding relationship of formula 2-2, we can obtain a continuous phase image of a cycle reprojected from the projector's perspective to the camera's perspective: , thus constructing a cycle phase consistency loss, as shown in Formula 3-3 (3-3) In the formula, the set represents a valid point, which is exactly the same as the valid point introduced in Section 2.2. Indicate point The absolute phase value of one cycle reprojected to the camera plane, Indicates the absolute phase value of the effective point in the camera plane over one cycle.
[0072] Finally, after adding two losses to constrain the depth map, Figure 5 As shown in the figure, the first row is a 64-cycle fringe image, the second row is a 64-cycle truncated phase image, the third row is a true depth map, the fourth row is a depth map generated using only the consistency loss supervision of the 64-cycle truncated phase, the fifth row is a depth map generated using only the consistency loss supervision of one cycle of continuous phase, and the sixth row is a depth map generated by jointly supervising the two losses.
[0073] S4. The results of 3D reconstruction using this method can generate a depth map close to the true value after supervision using two losses. Figure 6 The 3D reconstruction effect of the depth map generated by the proposed weakly supervised method is demonstrated.
[0074] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A weakly supervised method for estimating depth from truncated phase, characterized in that The following steps are involved: S1, data preprocessing, recovering the truncated phase image from the fringe image; S2, supervised depth map generation using consistency loss of truncated phase; S3, supervise the depth map generation using a one-cycle continuous phase consistency loss; S4. Generates a depth map close to the ground truth after supervision using two losses.
2. A weakly supervised method for estimating depth from truncated phase according to claim 1, characterized in that: The method comprises the following steps: according to the operation steps in S1, S101, first collect 64 cycles of three-step phase-shift fringe patterns from the FPP system , by adjusting the order of the three-step phase-shifted fringe images; S102. Calculate three different truncated phase images according to the formula .
3. A weakly supervised method for estimating depth from truncated phase according to claim 2, characterized in that: The following steps are included: According to the operation steps in S102, the formula In the formula, , , is a 64-cycle three-step phase-shift fringe pattern collected from the FPP system. is the calculated truncated phase map.
4. A weakly supervised method for estimating depth from truncated phase according to claim 3, characterized in that: The method comprises the following steps: according to the operation steps in S2, S201, the 64-cycle truncated phase under the projector viewing angle is fixed, and the 64-cycle sinusoidal fringe encoding is obtained using the formula; S202: Through the preset fringe coding, the 64-cycle truncated phase image under the projector viewing angle can be calculated. ; S203, input the obtained truncated phase image into the depth estimation network, and then an initial estimated depth map can be obtained. The resolution of the depth map is the same as the input truncated phase image. When the posture relationship between the camera and the projector and the depth D corresponding to each pixel coordinate i, j in the depth map are known, the depth map can be converted into a point cloud in three-dimensional space and reprojected onto the projector plane, so that the pixel coordinate value of each point in the depth map can be calculated. The corresponding pixel coordinate value under the projector's viewing angle ; S204, reprojecting the 64-cycle truncated phase image from the projector's perspective to the camera's perspective. Since the u and v pixel coordinate values obtained by reprojection are generally floating point numbers, the truncated phase image from the projector's perspective The pixel coordinates are integers, so bilinear interpolation is needed to obtain the corresponding pixel coordinate values. Finally, a 64-cycle truncated phase image under the camera's perspective can be obtained. ; S205, in getting Then, compare it with the input A supervisory signal is established to construct the truncated phase consistency loss.
5. A weakly supervised method for estimating depth from truncated phase according to claim 4, characterized in that: The following steps are involved: According to the operation steps in S201, the formula is: In the formula, , where h is the height of the image, represents the number of pixels in a fringe period, is the sequence number of the three-step phase shift, represents the pixel coordinates on the projector plane, Refers to the projector.
6. A weakly supervised method for estimating depth from truncated phase according to claim 5, characterized in that: The following steps are involved: According to the operation steps in S204, the The calculation formula is In the formula, Represents the pixel coordinate value from the camera's perspective. Indicates the pixel coordinate value under the projector's viewing angle, with the superscript Refers to the projector, Indicates the coordinates under the projector's viewing angle The cutoff phase value of Indicates the coordinates from the camera's perspective The cutoff phase value.
7. A weakly supervised method for estimating depth from truncated phase according to claim 6, characterized in that: The following steps are included: according to the operation steps in S205, the ; In the formula, represents the L1 norm loss of two truncated phase images, represents three interpolated truncated phase images from the camera perspective, and The three truncated phase images representing the input correspond one to one; Representatives in Point to the structural similarity of the two corresponding truncated phase images, , represents the proportionality coefficient of different losses, Represents the effective point, Represents the set of valid points.
8. A weakly supervised method for estimating depth from truncated phase according to claim 7, characterized in that: The following steps are included: According to the operation steps in S205, the selection of the effective point can be obtained by the formula: In the formula, the set represents the effective point obtained by setting the modulation threshold. represents a point in the set, represents the serial number of the truncated phase image, Indicates On the picture The truncated phase value of the point, express The flag to indicate whether the point is selected. 1 means it is selected, and 0 means it is not selected.
9. A weakly supervised method for estimating depth from truncated phase according to claim 8, characterized in that: The following steps are involved: According to the operation steps in S3, S301, first collect a period of fringe pattern from the FPP system ; S302, using a one-cycle three-step phase shift fringe pattern, a one-cycle truncated phase pattern can be obtained through a formula; S303, after obtaining a truncated phase diagram of one cycle, it can be converted into an absolute phase through a formula; S304, using a formula to calculate a one-cycle absolute phase diagram under the camera's viewing angle , the absolute phase of one cycle on the projector plane is known and is expressed as , the positional relationship between the projector and the camera is known. Through the corresponding relationship of the formula, we can get a continuous phase image of a cycle reprojected from the projector's perspective to the camera's perspective. , thereby constructing a cycle phase consistency loss; S301. Finally, two more losses are added to constrain the obtained depth map comparison.
10. A weakly supervised method for estimating depth from truncated phase according to claim 9, characterized in that: The following steps are involved: According to the operation steps in S302, the formula In the formula, represents a fringe pattern of one cycle, represents the pixel coordinate value on the fringe pattern, represents the serial number of the fringe pattern; According to the operation steps in S303, the formula In the formula, is the pixel coordinate The truncated phase value at is the pixel coordinate calculated according to the formula The absolute phase value at ; According to the operation steps in S304, the formula In the formula, the set represents the effective point, which is exactly the same as the effective point mentioned above. Indicate point The absolute phase value of one cycle reprojected to the camera plane, Indicates the absolute phase value of the effective point in the camera plane over one cycle.
