Single image camera parameter estimation method, system and apparatus based on cross-field geometric consistency

By constructing full-field and zoom-field views and utilizing a shared-weight image encoder and a dual-branch decoder, explicit cross-field consistency constraints are applied, solving the ill-conditioned problems in single-image camera parameter estimation and the destruction of geometric consistency by image editing, thus achieving high-precision and robust camera parameter estimation.

CN123123705APending Publication Date: 2026-11-17ZHEJIANG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611242854.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing single-image camera parameter estimation techniques suffer from problems such as ill-conditioned principal point estimation, disruption of geometric consistency by image editing operations, and lack of cross-view geometric constraints in open scenes, leading to a decrease in camera parameter estimation accuracy.

Method used

We construct a full-field view and a zoom-field view, jointly process these two views using a shared-weight image encoder and a dual-branch decoder, and explicitly apply geometric constraints to improve robustness by aligning the light beam directions in the overlapping areas through cross-field consistency constraints.

Benefits of technology

It significantly improves the robustness and stability of single-image camera parameter estimation, and can maintain high-precision camera parameter estimation when faced with edited images and image distortion, especially showing excellent anti-cropping robustness in the case of image cropping and scaling.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure FT_1
    Figure FT_1
  • Figure FT_2
    Figure FT_2
  • Figure FT_3
    Figure FT_3
Patent Text Reader

Abstract

This invention discloses a single-image camera parameter estimation method, system, and apparatus based on cross-field-of-view geometric consistency. Addressing the ill-conditioned nature of single-image principal point estimation and its susceptibility to image cropping and scaling interference, a full-field-of-view and a zoom-field-of-view view are constructed from a single input image. A shared-weight image encoder and a dual-branch decoder jointly process the two views. Based on known pixel correspondences, a cross-field-of-view consistency regularization loss is explicitly applied to force the predicted ray beam directions of corresponding pixels within overlapping regions to be consistent. After training, the model can stably predict ray beam directions and input them into an external solver to recover camera parameters. This invention significantly improves the accuracy of camera parameter estimation and the stability of principal point estimation under image editing operations such as cropping and scaling. It achieves zero-shot evaluation results superior to existing methods on multiple public datasets and is applicable to computer vision tasks such as 3D reconstruction, SLAM, and motion recovery.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, computer vision and image processing technology, and in particular to a method, system and apparatus for single-image camera parameter estimation based on cross-field geometric consistency. Background Technology

[0002] Single-image camera parameter estimation aims to recover the camera's parameter matrix (including focal length and principal point) from a single image. It forms the basis for downstream computer vision tasks such as 3D reconstruction, Simultaneous Localization and Mapping (SLAM), and motion recovery. Traditional calibration methods often rely on calibration boards, vanishing points, or multi-view geometry, limiting their applicability in open scenes. In recent years, deep learning-based methods have made progress by predicting intermediate geometric representations such as perspective, incident, or field of view, and then combining this with data augmentation and optimized solvers to calculate camera parameters.

[0003] However, existing single-image camera parameter estimation techniques still have the following drawbacks: The first point is the ill-conditioned problem of principal point estimation. Estimating the principal point at any location directly from a single view lacks sufficient geometric constraints. Traditional methods usually assume that the principal point is located at the center of the image, but in actual shooting, physical misalignment between the sensor and the lens or lens eccentricity often causes the principal point to deviate from the center, making this assumption invalid.

[0004] The second point is the disruption of geometric consistency caused by image editing operations. In data augmentation and real image processing workflows, the common cropping and scaling operations change the effective field of view and image planarization parameters, causing the principal point to shift significantly in the cropping coordinate system. Existing methods mainly rely on single-view prediction and random data augmentation to adapt to these transformations, but data augmentation can only expose the model to diverse inputs and cannot provide explicit geometric constraints to address underlying ill-conditioning.

[0005] Thirdly, there is a lack of cross-view geometric constraints. Although different cropped views of the same physical scene have different image plane parameterizations, the underlying camera parameters and geometric poses remain unchanged, and corresponding pixels share the same ray beam direction. Existing methods process each input image independently, failing to utilize this inherent ray beam direction invariance, which leads to a significant decrease in the accuracy of camera parameter estimation when dealing with heavily cropped images. Summary of the Invention

[0006] This invention addresses the aforementioned problems by proposing a single-image camera parameter estimation method, system, and apparatus based on cross-field-of-view geometric consistency. The concept involves constructing a pair of full-field-of-view views and a cropped, scaled zoom-field-of-view view from a single input image. These two views are jointly processed using a weighted image encoder and a dual-branch decoder with an interactive mechanism. Cross-field-of-view consistency constraints are explicitly applied to align the predicted ray beam directions of corresponding pixels within overlapping regions. Implicit geometric invariance is transformed into explicit output space constraints, supplementing the ill-conditioned principal point estimation with missing geometric constraints. This significantly improves the robustness of the single-image camera parameter estimation method in the face of edited images and image distortion, as well as the stability of principal point estimation.

[0007] In a first aspect, the present invention proposes a single-image camera parameter estimation method based on cross-field-of-view geometric consistency, the steps of which include: Step 1: Construct a cross-field-of-view paired image input dataset, which includes a training set and a test set: When constructing the training set, import the fictional images rendered by the virtual camera; when constructing the test set, import the publicly available real images; for each fictional image and publicly available real image, construct a full field of view and a zoom field of view, and record the transformation matrix from the full field of view to the zoom field of view. Based on the transformation matrix, establish the correspondence between the corresponding pixels in the overlapping area from the full field of view to the zoom field of view. The purpose of using a virtual camera is to enable the accurate calculation of the direction of light beams for intrinsic parameters in a virtual image.

[0008] The purpose of the above steps is to ensure that the virtual images and the publicly available real images do not overlap, thus ensuring the independence of training and testing; Step 2: Construct a camera parameter estimation model based on cross-field geometric consistency; The camera parameter estimation model includes a weighted image encoder, a dual-branch decoder, and a ray prediction head; A weighted image encoder receives raw images of the full-field view and zoom-field view and encodes them into a feature space; a dual-branch decoder receives the encoded image features and performs joint feature decoding on the image features of the full-field view and zoom-field view to obtain dual-view features; a ray prediction head receives the dual image features, predicts the pixel-by-pixel ray beam direction and inputs it into an external camera parameter solver; the camera parameter solver fits and solves for camera parameters based on the pixel-by-pixel ray beam direction. Step 3: Train the camera parameter estimation model using the training set. During training, the model parameters are jointly optimized using independent ray supervision loss for each field of view and cross-field of view consistency regularization loss. The cross-field consistency regularization loss forces the predicted ray beam directions of corresponding pixels within the overlapping area to be consistent, based on the pixel correspondence. Step 4: Obtain the final camera parameter estimation model and input it into the test set for camera parameter estimation.

[0009] More specifically, step 1 includes: Step 1.1: Import virtual images and expose real images; set the target image size and target aspect ratio, and extract the view that retains the largest effective area and matches the target aspect ratio from the single input image as the full field of view; Step 1.2: Within the effective area of ​​the full field of view, crop according to the preset cropping area ratio and random position, and scale the cropped image to the same output resolution as the full field of view to obtain the zoom field of view; Step 1.3: Record the cropping and scaling transformation matrix from the full field-of-view image plane to the zoom field-of-view image plane, and establish the pixel correspondence based on the transformation matrix.

[0010] More specifically, the structure of the camera parameter estimation model described in step 2 includes: a shared-weight image encoder initialized with pre-trained weights; a dual-branch decoder initialized with pre-trained weights; and a beam prediction head.

[0011] A weighted image encoder receives full-field and zoom-field views for feature extraction and outputs image features. A dual-branch decoder performs self-attention operations to establish the context of the full-field and zoom-field views, and then performs joint feature decoding on the features of the full-field and zoom-field views. A ray beam prediction head predicts dense pixel-by-pixel ray beam directions from the features decoded by the dual-branch decoder and transmits them to an external camera parameter solver. The camera parameter solver fits and solves for camera parameters from the ray beam directions.

[0012] More specifically, the independent ray supervision loss for each field of view is the error calculated by projecting the predicted ray beam direction and the actual ray beam direction onto the parameterized representation of the tangent plane based on the field of view.

[0013] More specifically, the cross-field consistency regularization loss includes: extracting the predicted ray beam direction of each pixel in the full field of view and the predicted ray beam direction of the corresponding pixel in the zoom field of view within the overlapping region according to the pixel correspondence, and calculating the cosine similarity error between the two.

[0014] Secondly, this invention proposes a single-image camera parameter estimation system based on cross-field geometric consistency, comprising: The cross-field-of-view paired image construction module extracts the full field of view from the input single image, randomly crops a piece of image within the effective area of ​​the full field of view and scales it to the same size as the full field of view to obtain the zoom field of view, and establishes the pixel correspondence between the overlapping areas of the full field of view and the zoom field of view. The feature extraction module uses a weighted image encoder to extract features from the full field-of-view and zoom field-of-view views to obtain image features. The dual-branch interactive decoding module performs self-attention operations to establish the context of the full field of view and the zoom field of view, and then performs joint feature decoding on the image features of the full field of view and the zoom field of view to obtain the decoded image features. The ray prediction module receives the decoded image features and predicts the direction of the ray beam pixel by pixel based on the decoded image features; The camera parameter solving module receives the direction of the light beam pixel by pixel and fits and solves the camera parameter matrix.

[0015] Thirdly, the present invention proposes a single-image camera parameter estimation device based on cross-field geometric consistency, comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the above method.

[0016] In a fourth aspect, the present invention proposes a computer-readable storage medium, characterized in that it stores a program thereon, which, when executed by a processor, implements the above-described method.

[0017] The innovation of this invention lies in: (1) A method for supplementing geometric constraints in principal point estimation based on explicit optimization of overlapping ray beams is proposed. In the scenario of single-view camera parameter estimation, this invention overcomes the limitation of existing methods based on single-view camera parameter prediction that rely solely on data augmentation to adapt to clipping transformations. It discovers and utilizes the physical invariance that pixels in overlapping regions of the same physical image share the same ray beam direction under different clipping and scaling fields of view. By explicitly applying cross-field geometric consistency constraints, it supplements the missing geometric constraints for ill-conditioned principal point estimation. The essential characteristic of this constraint mechanism is that the constraint is directly applied to the ray beam direction with clear physical geometric meaning, rather than through indirect implicit guidance at the feature layer; the goal of the constraint is to make the network output satisfy known geometric transformation relationships, rather than simply learning the mapping between input and output from a statistical data perspective. This enables the network to actively correct principal point prediction bias during training, fundamentally suppressing principal point drift caused by field reparameterization.

[0018] (2) A training framework for resisting image editing interference based on zoom field of view is proposed. Existing single-image camera parameter estimation methods generally experience a significant decrease in parameter estimation accuracy when faced with common image editing operations such as changes in input image size, cropping, and resolution adjustment. The reason for this is that existing methods only use original images of fixed size as input during the training phase, without introducing training data distribution that matches the target application scenario (such as the reality of images being arbitrarily cropped and scaled), resulting in the model lacking adaptability to changes in image plane parameterization. By constructing a zoomed-out field of view as an auxiliary training signal, the network is forced to explicitly learn the invariance of ray beam direction under changes in image plane parameterization. The introduction of this zoomed-out field of view has a dual effect: First, it constitutes a "local magnification" observation of the full field of view. The network needs to maintain a consistent geometric understanding at both the full field of view scale (global spatial structure perception) and the zoomed-out scale (local detail perception), thus forcing the model to learn scale-invariant geometric representations unaffected by scale changes. Second, different field of view views are constrained by the cross-field geometric consistency regularization loss of overlapping ray beams, enabling the network to establish a cross-scale geometric consistency mapping in the output space. This significantly enhances the model's resistance to image editing interference such as arbitrary cropping ratios and positions. Even if the effective field of view of the input image is significantly reduced due to heavy cropping, the network can still accurately recover camera parameters based on the scale-invariant geometric priors learned during training.

[0019] (3) A unified framework for ray representation and cross-field joint inference is proposed. This invention constructs a network structure that includes a shared-weight image encoder and a dual-branch decoder with self-attention and cross-attention mechanisms. It can jointly process the full-field view and zoom-field view of the same image, and simultaneously realize cross-field information interaction and geometric alignment at the feature level and the output ray space level, effectively improving the robustness of the model in camera parameter estimation under complex image editing operations.

[0020] The beneficial effects of this invention include: (1) Significant principal point estimation stability. This invention directly aligns the predicted ray beam directions of corresponding pixels in overlapping areas of different views through cross-view geometric consistency constraints, fundamentally alleviating the ill-conditioned problem of single-view principal point estimation. When faced with interference from different degrees of clipping and scaling, it can maintain extremely low relative principal point estimation error, which is significantly better than existing methods that rely on data augmentation.

[0021] (2) Excellent calibration accuracy and robustness against cropping. Under the zero-sample evaluation setting with no previously seen images, the present invention can achieve or exceed the parameter estimation accuracy of the best existing methods on uncropped images. As the image is cropped, the proportion of the cropped image is reduced to 0.6. The performance degradation of the present invention in core indicators such as angle error and field of view error is slower than that of existing methods, showing significantly stronger robustness against cropping than existing methods.

[0022] (3) Wide model applicability. The cross-view geometric consistency mechanism of the present invention applies constraints uniformly in the output space, without relying on a specific camera parameterization form. Therefore, it can bring consistent performance improvement under various camera settings such as pinhole model, radial distortion model and general distortion model, demonstrating strong model generalization ability.

[0023] (4) No additional computational burden during inference. The cross-field consistency constraint proposed in this invention only functions as a regularization term during the training phase and does not require computation during the inference phase. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the steps of the single-image camera parameter estimation method based on cross-field geometric consistency of the present invention.

[0025] Figure 2 This is a schematic diagram of the zoom field of view structure of the present invention.

[0026] Figure 3 This is a network structure diagram of the camera parameter estimation model of the present invention.

[0027] Figure 4 This is a structural diagram of the single-image camera parameter estimation system based on cross-field geometric consistency according to the present invention. Detailed Implementation

[0028] To facilitate understanding of this application, the specific implementation of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of protection of the present invention. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of this application.

[0029] Example 1

[0030] according to Figure 1 The steps of the single-image camera parameter estimation method based on cross-field geometric consistency are as follows: Step 1: Construct cross-field-of-view paired image input data.

[0031] In this embodiment, the Open-Pano synthetic dataset is used as the training set, and the LaMAR, MegaDepth, TartanAir, and Stanford2D3D datasets are used as the test sets. For each input image in the dataset, the following construction process is performed: The target image size is set to 102400 pixels, and the target aspect ratio is uniformly sampled from the interval [0.5, 2.0]. The view that retains the largest effective area and matches the target aspect ratio is extracted from the input image and used as the full field of view.

[0032] The construction of the zoom field of view is as follows: Figure 2 As shown, the cropping area ratio (relative to the effective area of ​​the full field of view) is set to be uniformly sampled from the interval [0.2, 0.6], and the cropping position is uniformly randomly sampled from all feasible positions. The cropped area is then scaled to the same pixel resolution as the full field of view to obtain the zoomed field of view.

[0033] During the cropping and scaling process described above, the homography transformation matrix A from the full-field-of-view image plane to the zoom-field-of-view image plane is recorded. Based on matrix A, the full-field-of-view pixel structure is established. To zoom field of view pixels Correspondence This determines the coordinate mapping of corresponding pixels within the overlapping area of ​​the two views; Step 2: Construct a camera parameter estimation model based on cross-field geometric consistency.

[0034] The model network structure constructed in this embodiment is as follows: Figure 3 As shown, it includes a shared-weight image encoder, a dual-branch decoder, and two ray prediction heads.

[0035] In this embodiment, the image encoder and dual-branch decoder with shared weights initialize the self-pre-trained DUSt3R model weights. The shared encoder and cross-branch decoder, natively designed by the DUSt3R model, are naturally suited for joint inference on paired inputs and can provide good cross-view geometric correspondence priors.

[0036] The weighted image encoder receives the full-field view and zoom-field view image pairs generated in step 1 and maps them to a common feature space. The dual-branch decoder performs joint feature decoding on the full-field and zoom-field features. Each branch first applies a self-attention mechanism to model the context information within its respective view, and then applies a cross-attention mechanism to exchange information between the two branches, realizing cross-field joint feature interaction.

[0037] The features output from the dual-branch decoder are fed into two separate ray prediction heads, targeting the entire field of view.

[0038] The image and zoom field of view respectively output dense pixel-by-pixel corresponding ray beam direction predictions. and To stabilize the monitoring process, the beam direction is parameterized by projecting it onto the field of view (FoV) tangent plane. The external camera parameter solver receives... and The camera parameters are then solved by fitting the direction of the light beams pixel by pixel.

[0039] Step 3: Optimize the network parameters of the model using the training data.

[0040] During the training phase, the pre-trained image encoder weights with shared weights are frozen, and only the parameters of the dual-branch decoder and ray prediction head are optimized. The training process employs joint optimization using independent ray supervision loss for each view and cross-view consistency regularization loss: Specifically, independent ray monitoring loss for each field of view. yes:

[0041] Where Ω represents the image domain. It is by predicting the light beam Obtained by projecting onto the tangent plane, and It is based on the corresponding real light beam Calculated in the same way.

[0042] Cross-field consistency regularization loss yes:

[0043] in This represents the set of valid correspondences in the overlapping regions of the full field-of-view view and the zoom field-of-view view.

[0044] The final training loss is a weighted sum of the losses mentioned above, and the overall formula is:

[0045] in, It is a manually adjusted cross-field geometric consistency regularization loss weight hyperparameter. In this embodiment, it has been experimentally verified that... Setting it to 0.2 achieves the optimal balance by maintaining stable cross-view prediction without compromising the flexibility of single-view supervision.

[0046] The loss is calculated using backpropagation, and the network parameters are iteratively updated using an optimizer.

[0047] Step 4: Use the trained model network to estimate camera parameters.

[0048] During the inference phase, given a single open-world image with camera parameters to be estimated, the network directly constructs a pair of full-field view / zoom-field view images as input and performs forward propagation to obtain the predicted dense pixel-by-pixel ray beam direction.

[0049] The predicted pixel-by-pixel ray beam direction is input into an external camera parameter solver. In this embodiment, the AnyClib solver is used. The solver can recover the parameter matrix of the target camera model, including the focal length. , Main point , and distortion coefficient .

[0050] A single-image camera parameter estimation system based on cross-field geometric consistency for performing the above method is as follows: Figure 4 As shown, it includes: The cross-field-of-view paired image construction module extracts the full field of view from the input single image, randomly crops a piece of image within the effective area of ​​the full field of view and scales it to the same size as the full field of view to obtain the zoom field of view, and establishes the pixel correspondence between the overlapping areas of the full field of view and the zoom field of view. The feature extraction module uses a weighted image encoder to extract features from the full field-of-view and zoom field-of-view views to obtain image features. The dual-branch interactive decoding module performs self-attention operations to establish the context of the full field of view and the zoom field of view. Then, it performs joint feature decoding on the image features of the full field of view and the zoom field of view through branch 1 and branch 2 to obtain the decoded image features. The ray prediction module receives the decoded image features and predicts the direction of the ray beam pixel by pixel based on the decoded image features; The camera parameter solving module receives the direction of the light beam pixel by pixel and fits and solves the camera parameter matrix.

[0051] Experimental verification and effect analysis: To verify the effectiveness of this invention, in this embodiment, the trained model underwent zero-shot evaluation on multiple publicly available benchmark datasets, including LaMAR, MegaDepth, TartanAir, and Stanford2D3D. The model was trained only on the Open-Pano synthetic dataset and was not fine-tuned on the test dataset. This embodiment compares the method of this invention with the following existing mainstream methods: 1. WildCamera 2. DiffCalib 3. GeoCalib 4. AnyCalib-pin 5. AnyCalib-edi. This embodiment uses the following metrics to evaluate the accuracy of camera parameter estimation: angular error, used to measure the overall accuracy of camera pose estimation; vertical field of view error and horizontal field of view error, used to measure the accuracy of focal length estimation; relative principal point error, used to measure the accuracy of principal point coordinate estimation. This metric is presented as a percentage, with lower values ​​indicating more accurate principal point estimation. During testing, this embodiment also cropped the original image, and tested the model's performance under three conditions: no cropping, light cropping (image retention ratio 0.8), and heavy cropping (image retention ratio 0.6), verifying the stability of the method of this invention when dealing with image cropping. The evaluation results are shown in the table below: Table 1. Zero-sample test results for camera parameter estimation without cropping.

[0052] Table 2. Zero-shot test results for camera parameter estimation with light cropping (image retention ratio 0.8).

[0053] Table 3. Zero-shot test results of camera parameter estimation with height cropping (image retention ratio 0.6).

[0054] Tables 1, 2, and 3 show that on the LaMAR, MegaDepth, TartanAir, and Stanford2D3D datasets, as the image cropping retention ratio decreases from 1.0 to 0.6, the relative principal point error of the method in this invention... It consistently maintains the lowest possible values, significantly outperforming methods such as AnyCalib-edit that rely solely on data augmentation. Even with a heavy cropping margin of 0.6, this invention leads existing methods in both angular error (AE) and vertical / horizontal field-of-view error, demonstrating that cross-field-of-view geometric consistency constraints effectively preserve geometric structure and exhibit extremely high stability in camera parameter estimation for cropped images.

[0055] The core of this embodiment lies in transforming the inherent ill-conditioned problem in single-image camera parameter estimation—especially the principal point estimation instability under image cropping and scaling operations—into a solution problem with explicit geometric constraints.

[0056] From the perspective of physical imaging, performing different cropping and scaling operations on the same physical image only changes the parametric form of the image plane (i.e., field-of-view reparameterization), and does not change the underlying physical camera parameters. Therefore, in the overlapping area of ​​the generated different field-of-view views, pixels with the same geometric position necessarily correspond to the same three-dimensional imaging ray passing through the optical center, that is, they share the same ray beam direction. This invention utilizes this "ray beam direction invariance" as a natural geometric constraint.

[0057] In its specific working mechanism, this invention first starts with a single input image and constructs a full-field view and a zoom-field view using known cropping and scaling transformations. Subsequently, these two views are input into a shared-weight image encoder and a dual-branch decoder constructed based on the weights of a pre-trained DUST3R model. The dual-branch decoder not only extracts the contextual features of each view through a self-attention mechanism, but more importantly, it performs cross-field information interaction and joint inference through a cross-attention mechanism, thereby establishing a correlation between the same physical imaging in different coordinate systems at the feature level.

[0058] During the network output phase, the model predicts the dense pixel-by-pixel ray beam directions for the two views. During training, this invention introduces a cross-view geometric consistency regularization loss, which, based on known pixel correspondences, explicitly forces the predicted ray beam directions of corresponding pixels within overlapping regions to align in 3D space. This mechanism essentially adds an extra geometric equation to the previously unconstrained ill-conditioned principal point estimation: the representation of the same ray in two different image coordinate systems mutually constrains each other, greatly limiting the degree of principal point drift. This allows the model to no longer passively adapt to cropping solely through data augmentation, but rather actively correct prediction biases through cross-view geometric consistency.

[0059] Ultimately, during the inference phase, although the cross-field-of-view geometric consistency regularization loss is not included in the calculation, the model, after being trained using the aforementioned mechanism, already possesses the ability to stably predict ray beam directions under field-of-view reparameterization. The predicted pixel-by-pixel ray beam directions are fed into a model-independent camera parameter solver, enabling high-precision and robust recovery of the parameter matrix of pinhole or distorted camera models. The entire process bridges the parameter differences between different fields of view through explicit geometric constraints, fundamentally improving the stability and accuracy of single-image camera parameter estimation in complex image editing scenarios.

Claims

1. A single-image camera parameter estimation method based on cross-field-of-view geometric consistency, characterized in that, Includes the following steps: Step 1: Construct a cross-field-of-view paired image input dataset, which includes a training set and a test set: When building the training set, import fictional images rendered by a virtual camera; when building the test set, import publicly available real images. For each fictional image and a publicly available real image, a full-field view and a zoom field view are constructed, and the transformation matrix from the full-field view to the zoom field view is recorded. Based on the transformation matrix, the corresponding pixel relationship of the overlapping area from the full-field view to the zoom field view is established. Step 2: Construct a camera parameter estimation model based on cross-field geometric consistency; The camera parameter estimation model includes a weighted image encoder, a dual-branch decoder, and a ray prediction head; A weighted image encoder receives raw images of the full field of view and the zoom field of view, and encodes them into the feature space; The dual-branch decoder receives the encoded image features and performs joint feature decoding on the image features of the full field of view and the zoom field of view to obtain dual-view features. A ray prediction head receives dual image features, predicts pixel-by-pixel ray beam directions, and inputs them into an external camera parameter solver; the camera parameter solver fits and solves for camera parameters based on the pixel-by-pixel ray beam directions. Step 3: Train the camera parameter estimation model using the training set. During training, the model parameters are jointly optimized using independent ray supervision loss for each field of view and cross-field of view consistency regularization loss. The cross-field consistency regularization loss forces the predicted ray beam directions of corresponding pixels within the overlapping area to be consistent, based on the pixel correspondence. Step 4: Obtain the final camera parameter estimation model and input it into the test set for camera parameter estimation.

2. The single-image camera parameter estimation method based on cross-field geometric consistency according to claim 1, characterized in that, Step 1 includes: Step 1.1: Import virtual images and expose real images; set the target image size and target aspect ratio, and extract the view that retains the largest effective area and matches the target aspect ratio from the single input image as the full field of view; Step 1.2: Within the effective area of ​​the full field of view, crop according to the preset cropping area ratio and random position, and scale the cropped image to the same output resolution as the full field of view to obtain the zoom field of view; Step 1.3: Record the cropping and scaling transformation matrix from the full field of view image plane to the zoom field of view image plane, and establish the pixel correspondence based on the transformation matrix.

3. The single-image camera parameter estimation method based on cross-field geometric consistency according to claim 1, characterized in that, The structure of the camera parameter estimation model described in step 2 includes: a shared-weight image encoder initialized with pre-trained weights; a dual-branch decoder initialized with pre-trained weights; and a beam prediction head. A weighted image encoder receives full-field and zoom-field views for feature extraction and outputs image features. A dual-branch decoder performs self-attention operations to establish the context of the full-field and zoom-field views, and then performs joint feature decoding on the features of the full-field and zoom-field views. A ray beam prediction head predicts dense pixel-by-pixel ray beam directions from the features decoded by the dual-branch decoder and transmits them to an external camera parameter solver. The camera parameter solver fits and solves for camera parameters from the ray beam directions.

4. The single-image camera parameter estimation method based on cross-field geometric consistency according to claim 1, characterized in that, The independent ray supervision loss for each field of view is the error calculated by projecting the predicted ray beam direction and the actual ray beam direction onto the parameterized representation of the tangent plane based on the field of view.

5. The single-image camera parameter estimation method based on cross-field geometric consistency according to claim 1, characterized in that, The cross-field consistency regularization loss includes: extracting the predicted ray beam direction of each pixel in the full field of view and the predicted ray beam direction of the corresponding pixel in the zoom field of view in the overlapping area according to the pixel correspondence, and calculating the cosine similarity error between the two.

6. A single-image camera parameter estimation system based on cross-field geometric consistency, used to implement the method of claim 1, characterized in that, include: The cross-field-of-view paired image construction module extracts the full field of view from the input single image, randomly crops a piece of image within the effective area of ​​the full field of view and scales it to the same size as the full field of view to obtain the zoom field of view, and establishes the pixel correspondence between the overlapping areas of the full field of view and the zoom field of view. The feature extraction module uses a weighted image encoder to extract features from the full field-of-view and zoom field-of-view views to obtain image features. The dual-branch interactive decoding module performs self-attention operations to establish the context of the full field of view and the zoom field of view, and then performs joint feature decoding on the image features of the full field of view and the zoom field of view to obtain the decoded image features. The ray prediction module receives the decoded image features and predicts the direction of the ray beam pixel by pixel based on the decoded image features; The camera parameter solving module receives the direction of the light beam pixel by pixel and fits and solves the camera parameter matrix.

7. A single-image camera parameter estimation device based on cross-field geometric consistency, characterized in that, The device includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the method of any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the method of any one of claims 1-6.