A frequency-aware flow aided underwater target pose estimation self-supervised method

Abstract

A frequency-aware flow-assisted underwater target pose estimation self-supervised method, comprising: performing frequency domain analysis on synthetic and real underwater RGB images, extracting amplitude and phase information, and performing frequency domain amplitude processing strategy to enhance the diversity and robustness of the data; secondly, the pre-trained pose and flow estimation network is used to obtain the initial pose and flow field, and the real image is preliminarily predicted based on the initial pose; further, based on the 3D model and the predicted initial pose, the changed synthetic image is rendered, and rich perspective changes are provided around the initial pose. Finally, combined with real and synthetic images, through dense 2D-2D correspondence and feature level alignment, iterative optimization of pose and optical flow is carried out, and shape constraint is used to enhance the self-supervised learning ability. The method effectively improves the accuracy and robustness of pose estimation, avoids the dependence on depth data, reduces the cost, and widens the potential of underwater application scenarios.

Description

Technical Field

[0001] This invention relates to object pose estimation, and in particular to a frequency-aware flow-assisted self-supervised method for underwater target pose estimation. Background Technology

[0002] 6D object pose estimation for underwater unmanned vehicles has wide applications, such as underwater vehicle tracking, underwater archaeology, marine resource development, and underwater intervention missions. Significant progress has been made in general-purpose 6D object pose estimation, particularly in indoor scenarios. However, due to the complexity of the underwater environment, these methods often fail to achieve satisfactory results when directly applied to underwater object pose estimation. Therefore, accurately estimating the 6D pose of objects in underwater environments remains a formidable challenge.

[0003] The challenges of underwater object pose estimation are mainly threefold. First, due to complex optical effects underwater, realistic underwater images typically exhibit greater degradation in terms of illumination conditions and sharpness, resulting in a larger domain gap between them and synthetic images than in typical scenes. Second, because near-infrared light attenuates faster underwater than visible light, ordinary depth cameras often fail to function. Furthermore, obtaining realistic pose annotations for underwater objects is significantly more costly due to the greater difficulty of deploying additional calibration equipment underwater and the lack of depth data. To address the challenges of annotating realistic underwater data, existing techniques primarily focus on synthetic underwater style data or self-supervised pose estimation. These techniques have made some progress, with representative works including:

[0004] (1) Synthetic underwater style data

[0005] To circumvent the challenge of annotating real underwater data, some existing works have focused on training on synthetic underwater-style data generated by Unreal Engine 4 (UE4), Unity3D, or Physically Based Rendering (PBR). However, due to the domain differences between synthetic and real images, networks trained on synthetic data often exhibit poor generalization ability on real-world data. To improve the accuracy of pose estimation, a common approach is to refine the initial prediction using rendering and comparison paradigms. While this approach is effective to some extent, it still cannot completely eliminate the significant domain differences between the synthetic and real domains.

[0006] (2) Self-supervised pose estimation

[0007] Self-supervised object pose estimation methods have demonstrated their effectiveness in mitigating the domain gap in general scenarios by leveraging unlabeled real images to enhance pre-trained networks. However, the lack of depth data in underwater environments renders most depth-dependent methods unsuitable. Hai et al. proposed integrating pose refinement and self-supervised learning through pseudo-flow consistency, eliminating the need for depth. However, these methods typically self-train by treating the network as a black box, i.e., only forcing alignment with the outputs of teacher and student networks, ignoring the influence of the feature space.

[0008] Therefore, the existing methods have the following shortcomings in practical applications:

[0009] (1) Differences between synthetic and real data: Although synthetic data can be generated in large quantities, the differences between it and real underwater data are still significant. These differences include lighting conditions, suspended matter, color decay, etc., which lead to poor performance of the trained model in practical applications.

[0010] (2) Robustness of self-supervised methods: Although self-supervised methods reduce the dependence on labeled data, they still face challenges in complex underwater environments. For example, changes in lighting conditions, interference from suspended objects, and image blurring can all affect the robustness and accuracy of the algorithm.

[0011] (3) Generalization performance of the algorithm: Due to the diversity of underwater environments, the algorithm needs to have strong generalization ability and perform well under different water conditions, depths, seasons, etc. This places higher demands on the algorithm.

[0012] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0013] The main objective of this invention is to solve the problems existing in the above-mentioned background technology and provide a frequency-sensing flow-assisted self-supervised method for underwater target pose estimation.

[0014] To achieve the above objectives, the present invention adopts the following technical solution:

[0015] In a first aspect of the present invention, a frequency-aware flow-assisted self-supervised method for underwater target pose estimation includes the following steps:

[0016] S1. Acquire synthetic RGB images and real underwater captured RGB images; perform frequency domain analysis on the images using frequency domain transformation methods to extract amplitude and phase information, and execute frequency domain amplitude processing strategies to achieve data augmentation;

[0017] S2. Using the synthetic image augmented by the data in step S1, pre-train a preliminary pose estimation network and a flow estimation network to obtain the initial pose and initial flow field of the image; use the preliminary pose estimation network to perform preliminary pose estimation on the real underwater image to obtain the initial pose prediction.

[0018] S3. Using the given 3D model of the target object and the initial pose predicted in step S2, render multiple synthetic images that vary around the initial pose.

[0019] S4. Based on the preliminary pose obtained in step S2 and the synthetic image generated in step S3, iterative optimization of pose and optical flow estimation is performed. This includes establishing a dense 2D-2D correspondence between the real and synthetic images and using optical flow to assist in image-level alignment; performing feature-level alignment in the high-dimensional latent space to extract and integrate high-level semantic information; and combining image-level and feature-level alignment with shape-constrained optical flow to perform self-supervised learning of the network.

[0020] In a second aspect of the invention, a computer program product includes a computer program that, when executed by a processor, implements the underwater target pose estimation self-supervised method.

[0021] The present invention has the following beneficial effects:

[0022] This invention proposes a frequency-aware flow-assisted self-supervised method for underwater target pose estimation. Through an innovative two-stage self-supervised framework, it effectively addresses several key issues in existing underwater pose estimation techniques. First, to address the domain gap problem between synthetic and real underwater data, this invention employs an FFT-based data augmentation strategy. By mixing amplitudes and discarding amplitudes, it introduces style priors from the unlabeled target domain, simultaneously enabling the network to learn domain-invariant features, significantly improving the model's adaptability to different underwater scenarios and the accuracy of pose estimation. Second, to enhance the algorithm's robustness, this invention proposes an optical flow-assisted multi-level alignment strategy, combining image-level and feature-level alignment. This achieves consistent forward and backward flow alignment at the image level and feature-level alignment in a high-dimensional latent space, effectively extracting and integrating high-level semantic information, improving the effectiveness of self-supervised learning, particularly in complex underwater environments. Furthermore, this invention's advantages also lie in its ability to estimate pose without relying on depth data, reducing data acquisition costs, and simplifying the training process through a self-supervised refinement strategy, shortening training time, and improving training efficiency. Meanwhile, the self-supervised learning method of this invention demonstrates high efficiency in complex environments, providing a high-precision, low-cost, and high-efficiency solution for fields such as underwater archaeology, marine resource development, and robot navigation, with broad market application prospects. In summary, this invention, through its innovative technical solution, not only improves the accuracy and robustness of underwater target pose estimation but also possesses cost-effectiveness and broad application potential.

[0023] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0024] Figure 1 This is an algorithmic framework diagram of the frequency-sensing flow-assisted self-supervised underwater target pose estimation method according to an embodiment of the present invention.

[0025] Figure 2 This is a flowchart illustrating the frequency domain sensing data enhancement strategy according to an embodiment of the present invention.

[0026] Figure 3 This is a flowchart illustrating the overall process of the self-supervised underwater target pose estimation method according to an embodiment of the present invention. Detailed Implementation

[0027] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and is not intended to limit the scope and application of the present invention.

[0028] This invention effectively solves the problems of accuracy and robustness in 6D object pose estimation in underwater environments. Existing methods perform poorly when faced with complex optical effects, lack of depth data, and high-cost data annotation in underwater environments. Specifically, existing methods have significant shortcomings in terms of domain gaps between synthetic and real data, changes in lighting conditions, interference from suspended objects, and image blurring, resulting in poor generalization ability and robustness in practical applications. This invention introduces a data augmentation method based on frequency domain transformation methods such as Fast Fourier Transform (FFT) and a multi-level flow-assisted consistency strategy, thereby significantly improving the accuracy and robustness of 6D pose estimation in underwater environments. Specifically, it includes: an FFT-based data augmentation strategy: by mixing amplitude information from synthetic and real images and randomly discarding some amplitude information, it helps the network learn domain-invariant features. An optical flow-assisted multi-level alignment strategy: by aligning at the image level and feature level, it enhances the self-supervised learning effect of the network.

[0029] In summary, the working process of this invention mainly includes a pre-training stage and a self-supervised refinement stage:

[0030] Pre-training phase:

[0031] Data augmentation: The amplitude information of the synthetic and real images is mixed using FFT, while some amplitude information is randomly discarded. This augmentation strategy allows the network to undergo initial pre-training on synthetic data.

[0032] Preliminary pose estimation: A pre-trained network is used to perform preliminary pose estimation on real underwater images.

[0033] Self-monitoring refinement stage:

[0034] A flow-assisted self-supervised framework is proposed. An object pose estimator is built upon the flow-assisted direct pose regression network for iterative optimization of pose and optical flow. During training, a flow-assisted multi-level alignment strategy (including image-level and feature-level alignment) is employed to enhance the performance of the self-supervised network.

[0035] Image-level alignment: An optical flow-assisted image-level alignment strategy is introduced, which is refined through a consistent forward and backward flow alignment method.

[0036] Feature-level alignment: Performing feature-level alignment in a high-dimensional latent space to extract and integrate high-level semantic information.

[0037] See Figures 1 to 3 This invention provides a frequency-aware flow-assisted self-supervised method for underwater target pose estimation, comprising the following steps:

[0038] S1. Data Preprocessing and Augmentation:

[0039] Collect synthetic RGB images and real underwater captured RGB images;

[0040] The image is analyzed in the frequency domain using frequency domain transformation methods such as Fast Fourier Transform (FFT) to extract amplitude and phase information, and a frequency domain amplitude processing strategy is executed to achieve data augmentation.

[0041] S2. Pre-trained network model:

[0042] The initial pose estimation network and flow estimation network are pre-trained using the synthetic image with data augmentation in step S1 to obtain the initial pose P0 and initial flow field of the image.

[0043] The preliminary pose estimation network is used to perform preliminary pose estimation on real underwater images to obtain initial pose prediction.

[0044] S3. Generate the composite image:

[0045] Using a given 3D model of the target object and the initial pose P0 predicted in step S2, render N synthetic images that vary around the initial pose P0.

[0046] S4. Self-supervised refinement:

[0047] Based on the initial pose P0 obtained in step S2 and the synthetic image generated in step S3, iterative optimization of pose and optical flow estimation is performed, including: image-level alignment: establishing a dense 2D-2D correspondence between the real image and the synthetic image generated in step S3, and using optical flow to assist in image-level alignment; feature-level alignment: performing feature-level alignment in a high-dimensional latent space to extract and integrate high-level semantic information; multi-level flow-assisted consistency strategy: combining image-level and feature-level alignment, using shape-constrained optical flow to enhance the self-supervised learning effect of the network.

[0048] In feature-level alignment, a multi-scale feature fusion strategy can be employed to extract more detailed information by fusing features at different scales. Simultaneously, multi-scale feature alignment operations can also be performed within feature-level alignment to ensure consistency in feature representations across different scales.

[0049] In a preferred embodiment, in step S1, the frequency domain amplitude processing strategy includes mixing and discarding amplitudes, or replacing amplitudes. The mixing and discarding of amplitudes includes: introducing style priors of the target domain by mixing phase information of the synthesized image with amplitude information of the real image, and randomly discarding amplitude information to enable the network to learn domain-invariant features.

[0050] In a preferred embodiment, in step S2, a preliminary pose estimation network and flow estimation network are pre-trained using the synthetic image augmented in step S1 to obtain the initial pose P0 and initial flow field of the image. Based on these two preliminary networks, the teacher network and student network for the self-supervised learning stage are initialized. The teacher network is initialized using the parameters of the pre-trained pose estimation network and flow estimation network, and the network structure or parameters are fine-tuned to generate high-quality pseudo-labels to guide the training of the student network. The student network is initialized based on the parameters of the pre-trained pose estimation network and flow estimation network, and will be iteratively optimized independently of the teacher network during the self-supervised learning process. In step S4, during the iteration process, the teacher network and the student network work collaboratively. The teacher network generates pseudo-labels to assist the training of the student network and guides the student network's predictions to be consistent with the teacher network's output by providing consistency feedback. The teacher network generates pseudo-labels as self-supervised signals to guide the training direction of the student network and provides auxiliary and consistency feedback to ensure that the student network's predictions are consistent with the high-accuracy output of the teacher network, thereby improving the accuracy and robustness of pose estimation.

[0051] In a preferred embodiment, step S1 specifically includes: S11. Performing Fast Fourier Transform (FFT) on the collected synthetic RGB images and real underwater captured RGB images to convert the spatial domain images to a frequency domain representation; S12. Separating the amplitude and phase information of the images from the FFT results, where amplitude information reflects the brightness and contrast of the image, while phase information contains the contour and edge features of the image; S13. Based on the amplitude and phase information, performing an amplitude mixing operation on the images to combine the amplitude information of the synthetic image with the amplitude information of the real image to introduce style features of the target domain; S14. Implementing an amplitude discarding strategy, randomly selecting a certain proportion of amplitude information to discard, prompting the network to focus on learning domain-invariant phase information, and enhancing the model's ability to... Generalization ability across different image domains; S15. Perform linear interpolation on amplitude information, using amplitude information from two images randomly sampled from a synthetic dataset and a real-world dataset, and generate a new amplitude spectrum through weighted averaging to introduce style prior; S16. Determine which amplitude information is retained and which is discarded based on a specific discard threshold to control the degree of amplitude spectrum discarding; S17. Combine the transformed amplitude information and the original phase information, reconstruct the image through inverse Fourier transform to generate an enhanced image for training, which integrates the style features of the synthetic image and the contour information of the real image; Repeat the above steps to generate an enhanced image set for all images used for training, which will be used for subsequent network pre-training and self-supervised refinement steps.

[0052] In a preferred embodiment, step S2 specifically includes: S21. Initializing the training process of the pose estimation network and the flow estimation network using the synthetic image obtained after data augmentation; S22. Predicting the pose of the real underwater image using the initial network to obtain the initial pose P0; S23. Based on the known ground pose... The corresponding ground flow was calculated. Then, based on the predicted initial pose P0, the predicted flow is calculated; S24. Supervised learning is used, combined with the ground real-time flow. S25. Calculate the flow loss L_flow by predicting the flow fs; S25. Employ supervised learning, combining the actual pose... S26. Calculate the pose loss L_pose based on the predicted pose Ps; S27. Determine the weight ratio of the flow loss and pose loss, and set the weight γ0 of the flow loss, for example, to 0.1; S28. Combine the flow loss L_flow and pose loss L_pose according to the determined weight ratio to form the total supervised loss L_sup; S29. Use the total supervised loss L_sup to guide the training of the network, update the network parameters through the backpropagation algorithm, and optimize the prediction accuracy of pose and flow; S20. Repeat steps S22 to S28 for multiple rounds of iterative training until the network achieves the expected accuracy in predicting flow and object pose; During the training process, continuously evaluate the network performance to ensure that the model can effectively learn from the augmented synthetic data and generalize to the pose estimation task of real underwater images.

[0053] In a preferred embodiment, step S4 specifically includes: S41. Establishing 2D-2D correspondences: Establishing dense pixel-level correspondences between the real image and each synthesized image, providing a basis for pose fine-tuning. S42. Shape constraint flow prediction: The network learns to predict shape constraint flows, that is, predicting the motion of pixels under continuous pose iterations based on the correspondences. S43. Pose iterative update: Combining the predicted shape constraint flows and camera parameters, iteratively updating the pose estimation to gradually refine the pose prediction. S44. 2D-3D correspondence calculation: In each iteration, calculating the correspondence between visible object points in the input image and 3D model points. S45. Pose guidance flow generation: Using the calculated 2D-3D correspondences, generating a pose guidance flow to guide further iterative optimization of the pose. S46. Decoupled pose estimation: Decoupling the rotation and translation components and estimating them separately to simplify the pose optimization problem and improve the accuracy of the estimation. S47. Continuous Rotation Representation: A continuous rotation representation method is adopted to avoid the use of quaternions or Euler angles, so that the neural network can directly regress the pose parameters. Repeat steps S42 to S47 for multiple iterations until the pose estimation reaches the predetermined accuracy or the upper limit of the number of iterations. The final pose estimation results can be verified and evaluated to ensure that their accuracy and reliability meet the requirements of the self-supervised framework.

[0054] In a preferred embodiment, step S4, the image-level alignment specifically includes the following operations: Predicting the flow field: The student network predicts the flow field between pixels in the real and synthesized images to obtain the student network's predicted flow field. Backward optical flow consistency: Based on the student network's predicted flow field, the real image is backward-warped and compared with the real image after pseudo-flow field transformation generated by the teacher network. Photometric consistency is used as a metric to calculate the backward optical flow consistency loss. Forward optical flow consistency: The consistency of the forward flow predicted by the student network on the warped rendering mask is evaluated. The forward optical flow consistency loss is calculated by comparing the forward warping transformation results of the predicted flow field and the pseudo-flow field. Optical flow self-supervision: The teacher network evaluates the consistency of the predicted flow between the synthesized image and the real image. Reliable pseudo-labels are selected using standard deviation to supervise the student network's flow field prediction. Calculating the image-level alignment loss: The backward optical flow consistency loss, the forward optical flow consistency loss, and the optical flow self-supervision loss are weighted and summed to obtain the total image-level alignment loss. The weights of each loss are adjusted according to the experimental settings. Optimize network parameters: Based on the image-level alignment loss, update the parameters of the student network using the backpropagation algorithm to optimize the accuracy of flow field prediction. Repeat the above steps for multiple iterations until the flow field prediction reaches the predetermined accuracy or the upper limit of the number of iterations, thus achieving refined image-level alignment.

[0055] In a preferred embodiment, step S4, the feature-level alignment specifically includes the following operations: Feature extraction: Using a feature encoder of a pose network, feature representations are extracted from real underwater images and synthetic images respectively, while ensuring that the feature extractor is robust to changes in the appearance of the underwater environment. Feature warping: Based on the flow field predicted by the student network, the features of the real image are warped to generate warped feature representations, taking into account the impact of environmental changes in this process. Feature-level loss calculation: The difference between the warped feature representations and the features of the synthetic image is calculated, and a metric that reflects feature similarity and adapts to environmental changes is used to define the feature-level loss. Feature-level alignment optimization: The feature-level loss is used to guide the training of the pose network. The network parameters are updated through a backpropagation algorithm to optimize the feature encoder to extract more aligned feature representations, emphasizing robustness to environmental changes in this process. The above steps are repeated for multiple iterations, with the robustness of the feature-level alignment evaluated in each iteration, until the alignment of the feature representations reaches a predetermined accuracy or an upper limit of the number of iterations, achieving fine-grained feature-level alignment and ensuring robustness to changes in the appearance of the underwater environment throughout the process.

[0056] In a preferred embodiment, step S4 introduces pose loss and total self-supervised loss during the self-supervised learning process, including: Pose transformation: Transforming the uniformly sampled 3D point set on the 3D model surface using the pseudo-labeled pose P^tea and the predicted pose P^stu to obtain the corresponding 2D point positions; Point matching loss calculation: Calculating the pose loss L_pose based on the transformed 2D point positions. This loss reflects the difference between the pseudo-labeled pose and the predicted pose, and the L1 norm is used to measure the accuracy of point matching. Pose optimization: Utilizing the pose loss to guide the optimization of pose parameters. By adjusting the pose parameters, the point matching loss is minimized, thereby improving the accuracy of pose estimation; Comprehensive loss definition: Constructing the total self-supervised loss L_self, which is a weighted sum of pose loss, image-level loss, and feature-level loss. Assigning associated weights to the image-level loss and feature-level loss to balance the impact of different losses on the training process. Network parameter update: Updating the network parameters using the backpropagation algorithm based on the calculated total self-supervised loss to optimize the performance of the pose estimation network. Repeat the iterative training process multiple times until the network reaches the predetermined accuracy or the maximum number of iterations for the pose estimation task. At the end of each iteration cycle, evaluate the accuracy of the pose estimation and the overall performance of the network to ensure that the pose loss and total self-supervised loss are minimized, while the network remains robust to various conditions.

[0057] The following describes specific embodiments of the present invention.

[0058] To address the aforementioned technical problems, this invention proposes a two-stage self-supervised 6D pose estimation framework, aiming to improve an object pose estimator initially trained on synthetic data and adapt it to real-world underwater data. This invention focuses on RGB-based 6D pose estimation without requiring additional sources of supervision. In the pre-training stage, inspired by frequency domain image analysis, a preferred embodiment introduces a data augmentation method based on Fast Fourier Transform (FFT), involving amplitude mixing and amplitude discarding. Specifically, by mixing amplitude information from synthetic and real images, a style prior from the unlabeled target domain is introduced. Simultaneously, randomly discarding amplitudes helps the network learn domain-invariant features. In the subsequent self-supervised refinement stage, multi-level flow-assisted consistency is introduced, performing alignment at both the image and feature levels, thereby enhancing network performance more effectively. Focusing solely on image-level alignment may overlook latent representations crucial for semantic and geometric consistency. Therefore, establishing feature-level alignment within a high-dimensional latent space, rather than just image-level alignment, enables more effective extraction and integration of high-level semantic information, making the self-supervision of this invention more effective.

[0059] (1) FFT-based augmentation strategy

[0060] Considering that phase contains contour and edge information and is insensitive to domain shift, while amplitude reveals image style, this invention proposes an FFT-based enhancement strategy that involves amplitude blending and amplitude discarding, rather than directly training on synthetic images. Specifically, by blending amplitude information from synthetic and real images, style priors from the unlabeled target domain are effectively introduced. Simultaneously, randomly discarding amplitudes helps the network learn domain-invariant features. This strategy effectively mitigates the domain differences between synthetic and real images, improves overall pose accuracy, and enhances the network's adaptability to different underwater scenes.

[0061] (2) Optical flow-assisted multi-level alignment strategy

[0062] This invention proposes an optical flow-assisted multi-level alignment strategy to promote pose refinement and improve the effectiveness of self-supervision. It mainly includes image-level alignment and feature-level alignment.

[0063] Image-level alignment: A forward and backward consistent flow alignment method is proposed, which simultaneously refines the flow field prediction by distorting the consistency of the forward flow on the rendering mask and the photometric consistency of the backward distorted real image, thereby enhancing network performance.

[0064] Feature-level alignment: Focusing solely on image-level alignment may overlook latent representations crucial for semantic and geometric consistency. Therefore, establishing feature-level alignment within a high-dimensional latent space, rather than just image-level alignment, enables more effective extraction and integration of high-level semantic information, making the self-supervision of this invention more effective.

[0065] Therefore, this invention proposes a frequency-aware flow-assisted self-supervised method for underwater target pose estimation, and completes the algorithm implementation and verification of the technical solution. The algorithm framework of this invention is as follows: Figure 1 As shown, given a 3D CAD model of an object, a set of synthetic RGB images with 6D pose annotations, and a set of unlabeled real RGB images captured underwater, the goal is to obtain accurate object poses from the real underwater RGB images. The working framework of this embodiment comprises two parts: a pre-training phase and a self-supervised phase. In the pre-training phase, unlike previous approaches that relied solely on synthetic data for training, a frequency-based data augmentation method is introduced to train the network model using the augmented data. In the self-supervised phase, shape-constrained optical flow is employed, and various alignment strategies are proposed to facilitate pose refinement. This framework is implemented based on a differentiable learning pipeline.

[0066] The following section details the processing flow and algorithm examples for the pre-training and self-supervised phases.

[0067] (1) Pre-training phase

[0068] In the pre-training phase, labeled data is used to train the network to iteratively predict flow and object pose. First, the initial pose P0 of the image is obtained using existing 6D pose estimation methods (such as GDR-Net and DeepIM). Then, a data augmentation strategy based on frequency domain transformation is proposed to encourage the network to not only acquire domain-invariant features but also introduce style priors related to the target domain. The augmented labeled data is then used to train the network.

[0069] 1) FFT-based augmentation strategies

[0070] Given an underwater image x, the phase and amplitude information are obtained using the FFT algorithm, as shown in the following formula:

[0071]

[0072] Where R(x) and I(x) are respectively The real and imaginary parts. Therefore, This represents the inverse Fourier transform. For example... Figure 2 As shown, amplitude and phase components are first extracted, and then the image is reconstructed based on each component. Notably, the image reconstructed by phase alone displays sharp contours, allowing for fast and direct object recognition. However, object recognition by amplitude-only reconstruction is not feasible. The study found that phase, which contains contour and edge information, is insensitive to domain shift, while amplitude reveals image style. Based on this property, the amplitude spectrum is randomly discarded to facilitate the network learning domain-invariant phase information. On the other hand, linear interpolation is performed on the amplitude spectra of two images from the source and target domains to generate new amplitude spectra, aiming to introduce style priors from the target domain. Specifically, given two images x randomly sampled from a synthetic dataset and a real-world dataset... s and x r The transformed amplitude information is defined as follows:

[0073]

[0074] Among them, the hyperparameter δ0∈(0,1) controls the discarding of the amplitude spectrum, and the hyperparameter α~U(0,β). The parameter β reflects the enhancement intensity.

[0075] Finally, the transformed amplitude information and the original phase information are reconstructed to generate an enhanced image x for training. s→r :

[0076]

[0077] 2) Pre-trained network

[0078] During the pre-training phase, the network is trained using augmented labeled data to iteratively predict flow and object poses. To supervise these predictions in each iteration, the network is first trained based on pose P0 and the ground-based real-world pose. Obtain ground live stream Then calculate the supervision loss. The supervised loss during the pre-training phase consists of the flow loss. and pose loss The composition, formula is as follows:

[0079]

[0080] in Is it related to the prediction result (f) s ,P s The corresponding truth label. In one example, the loss weight γ0 is set to 0.1. Calculate The results are obtained by training on synthetic data enhanced using the proposed FFT-based strategy.

[0081] (2) Self-monitoring stage

[0082] This invention proposes a flow-assisted self-supervised framework. An object pose estimator is built upon a flow-assisted direct pose regression network for iterative optimization of pose and optical flow.

[0083] Specifically, given the input real image I r Given the initial pose P0 of the target object, its 3D model is used first. To render N composite images surrounding P0 For stream-assisted object pose estimators, a dense 2D-2D correspondence is established between the real and synthetic images. Fine-tuning the pose P0 has been shown to improve the performance of self-supervised frameworks. Given in I r , The two pixel positions x in the matching t and x s The proposal suggests having the network predict shape-constrained flows.

[0084]

[0085] Among them, f s→t Represents a dense 2D-2D flow vector, where S is the iteration number and P is the flow vector. j Let M be the predicted pose at iteration j, and K be the matrix of camera intrinsic parameters. i It is the object-visible mask of the observed RGB image. To generate the pose-guided flow f s→t In each iteration, the 2D-3D correspondence between the input image and the points of the 3D object is calculated. Specifically, the points p of the 3D mesh model... i To 2D position u i =(u i ,v iThe correspondence between {p} i →u i} can be obtained in the following ways:

[0086]

[0087] Where λ i It is a scaling factor. R j and t j It is posture P j The rotational and translational components are then estimated. and This can be obtained given the initial pose P0. Then, the decoupled relative pose [R] is predicted and iteratively refined based on the initial pose. Δ |t Δ By decoupling the estimation of rotation from the estimation of translation, the significant error associated with the complex SE(3) space is effectively mitigated. Instead of using quaternions or Euler angles, a continuous 6-dimensional rotation representation is used to facilitate direct regression of pose in the neural network.

[0088] 1) Image-level alignment in self-supervised processes

[0089] For image-level alignment, the flow field f is based on the student network prediction. stu The pseudo-flow field f of the teacher network tea A learning objective was established, which includes forward-backward flow consistency and self-supervision of optical flow.

[0090] Backward optical flow consistency:

[0091] First, the goal of this invention is to utilize the predicted flow field f stu The real image is distorted backwards and passed through the pseudo-flow field f tea Alignment of the converted real images. Appearance similarity is measured using a widely adopted metric, photometric consistency. The photometric consistency of this invention... Real images distorted backward and Definition of differences between them:

[0092]

[0093] in ρ represents the back-torsion transformation based on the flow field, and calculates the photometric uniformity loss (e.g., Census loss). For simplicity, only the optical flow at 3D points on the visible surface is calculated.

[0094] Forward optical flow consistency:

[0095] The consistency of the forward flow on the warped rendering mask is evaluated as follows:

[0096]

[0097] Where ω represents the forward twist transformation based on the flow field, M s It is a mask for rendering images.

[0098] Optical flow self-supervision:

[0099] In addition, using a pseudo-flow field f tea To supervise the flow field f predicted by the student network stu To find reliable predictions for pseudo-labels, the teacher network of this invention evaluates the consistency of the prediction flow between N synthetic images and their corresponding real images using standard deviation. Specifically, the standard deviation is calculated as follows:

[0100]

[0101] Among them, f ij It is the prediction flow of the teacher network in the j-th iteration. This represents the pixel location in the real image. Pixels in these predictions are considered pseudo-labels only if their standard deviation is below a specified threshold σ0. tea and calculate To monitor the prediction of student networks stu For simplicity, σ0 = 1 is used in the experiment.

[0102] In summary, image-level alignment is calculated by the weighted sum of the above three factors:

[0103]

[0104] Here, γ1 and γ2 are the associated loss weights.

[0105] 2) Feature-level alignment in self-supervised processes

[0106] Compared to optical flow estimation for indoor RGB images, significant differences exist between real and synthetic images in underwater environments. These differences are primarily visual and stem from environmental variations such as changes in color, lighting, and sharpness. A feature-level error is introduced, which exhibits robustness to environmental variations, rather than relying solely on image-level errors. The pose network's feature encoder extracts features F from both real and synthetic underwater images. r and F s The proposed feature-level loss Defined by f stu Twisted F r and F s The weighted average of the dissimilarity of features between them.

[0107] 3) Pose loss during self-monitoring

[0108] In direct pose estimation, the choice of loss function is crucial for pose optimization. Point matching loss is used to align the poses P and P respectively. tea =[R tea |t tea ] and predicted pose P stu =[R stu |t stu The transformation of 3D points is calculated as follows:

[0109]

[0110] Where, p i These are 3D points uniformly sampled from the surface of the object model, where T represents the total number of points (1000 points were selected in the experiment).

[0111] 4) Total self-monitoring loss

[0112] The losses of self-monitoring can be summarized as follows:

[0113]

[0114] Here, γ3 and γ4 are the associated loss weights.

[0115] To demonstrate the effectiveness of this invention, several examples were designed to showcase its application effects under different circumstances.

[0116] Variation Example 1:

[0117] Specific steps:

[0118] (1) FFT-based data augmentation: The idea of ​​amplitude substitution is introduced, and the phase information of the synthetic image and the amplitude information of the real image are extracted by FFT and then mixed to reconstruct the augmented image. Through this augmentation strategy, the training network is initially pre-trained on the augmented synthetic data.

[0119] (2) The self-supervised phase is consistent with the best implementation.

[0120] Effect:

[0121] Directly replacing the amplitude information in the synthesized image may introduce color differences. The image after amplitude blending exhibits a more uniform appearance compared to the image with replaced amplitude. Experiments show that, before and after self-supervision, the network trained with data generated using the enhancement strategy of this invention performs better than the amplitude replacement strategy.

[0122] Variation Example 2:

[0123] Specific steps:

[0124] (1) Data Augmentation: Synthetic data generated based on PBR was used for training, and a new data augmentation method—frequency domain perturbation—was introduced. By perturbing the synthetic image in the frequency domain, the illumination changes and suspended object interference in the real underwater environment were simulated.

[0125] (2) Preliminary pose estimation: Preliminary pose estimation is performed on real underwater images using a pre-trained network.

[0126] (3) Self-supervised thinning: A new self-supervised thinning strategy is introduced—multi-scale alignment. Based on image-level alignment, multi-scale feature alignment is added. By performing feature-level alignment at different scales, the robustness and generalization ability of the network are improved.

[0127] Effect:

[0128] This deformation scheme effectively improves the accuracy and robustness of pose estimation through frequency domain perturbation and multi-scale alignment strategies. However, the performance of the algorithm still needs improvement when dealing with complex dynamic scenes.

[0129] The main innovations and advantages of this invention are:

[0130] (1) Data augmentation strategy based on FFT:

[0131] - By mixing amplitude information from synthetic and real images, a style prior for the unlabeled target domain is introduced.

[0132] - Randomly discarding amplitude information helps the network learn domain-invariant features.

[0133] (2) Optical flow-assisted multi-level alignment strategy:

[0134] - Image-level alignment: Enhance the network's image-level alignment capability through a consistent forward and backward flow alignment method.

[0135] - Feature-level alignment: Perform feature-level alignment in a high-dimensional latent space to extract and integrate high-level semantic information.

[0136] - Self-supervised refinement: Combining image-level and feature-level alignment improves the effectiveness of self-supervised learning.

[0137] The significant advantages of this invention are:

[0138] Mitigating the domain gap: FFT-based data augmentation strategies effectively mitigate the domain gap between synthetic and real data by mixing amplitude information from synthetic and real images and employing amplitude drop-off strategies.

[0139] Enhanced robustness: The optical flow-assisted multi-level alignment strategy performs alignment at both the image and feature levels, improving the algorithm's robustness in complex underwater environments.

[0140] Improved generalization ability: By performing feature-level alignment in a high-dimensional latent space, the algorithm's generalization ability under different water conditions, depths, seasons, etc., is enhanced.

[0141] Previous research has mainly focused on the generation of synthetic data and self-supervised learning, but has failed to fully consider the combination of optical flow-assisted and FFT data augmentation strategies. Furthermore, existing methods often neglect the importance of feature-level alignment. This invention, however, significantly improves the performance of self-supervised learning through a multi-level flow-assisted consistency strategy. In the self-supervised refinement stage, image-level and feature-level alignment is used to enhance the effectiveness of self-supervised learning.

[0142] Other embodiments

[0143] Besides FFT, other frequency domain transformation methods, such as Discrete Cosine Transform (DCT) or Wavelet Transform, can be explored to further enhance the diversity and robustness of the data. Regardless of the frequency domain transformation method used, the core idea is to introduce style priors from the unlabeled target domain while learning invariant information from the target domain.

[0144] In addition to RGB images, multimodal fusion can also combine sonar images, laser scanning data, or other sensor data to perform multimodal data fusion and improve the accuracy of pose estimation.

[0145] Optical flow-assisted strategies can introduce more complex optical flow calculation methods, such as deep learning-based optical flow estimation, to further improve the accuracy of image-level alignment. Regardless of the optical flow calculation method used, the core idea is to enhance the self-supervised learning effect of the network through multi-level alignment strategies.

[0146] By introducing an adaptive feature alignment strategy, the feature alignment parameters are dynamically adjusted according to different environmental conditions, thereby improving the generalization ability of the algorithm.

[0147] In feature-level alignment, a multi-scale feature fusion strategy is introduced to extract more detailed information by performing feature-level alignment at different scales.

[0148] This invention can be applied to the following technical scenarios:

[0149] Self-driving cars:

[0150] Object detection and pose estimation in complex urban environments help autonomous vehicles more accurately identify and locate surrounding objects, thereby improving driving safety.

[0151] Robot navigation and operation:

[0152] In complex industrial and home environments, robots can utilize this technology for accurate object recognition and pose estimation, thereby enabling more efficient navigation and manipulation tasks.

[0153] Augmented Reality (AR) and Virtual Reality (VR):

[0154] In AR and VR applications, accurate object pose estimation can help achieve more realistic interaction between virtual objects and the real environment, improving the user experience.

[0155] Medical image analysis:

[0156] In medical image analysis, especially in surgical navigation and robot-assisted surgery, accurate 3D pose estimation can help doctors operate more accurately and improve the success rate of surgery.

[0157] Drone navigation:

[0158] When drones fly in complex environments, this technology can be used for accurate object recognition and pose estimation, thereby achieving more stable flight and obstacle avoidance.

[0159] Ocean exploration and research:

[0160] In marine exploration and research, underwater robots can utilize this technology for precise object identification and pose estimation, thereby enabling more effective exploration and research of marine resources.

[0161] Industrial Automation:

[0162] In industrial automation, especially in assembly lines and quality inspection, robots can use this technology for accurate object recognition and pose estimation, thereby improving production efficiency and product quality.

[0163] Agricultural automation:

[0164] In agricultural automation, robots can use this technology for precise object recognition and pose estimation, thereby enabling automated tasks such as sowing, harvesting, and monitoring.

[0165] Logistics and warehousing management:

[0166] In logistics and warehousing management, this technology can be used to automate object recognition and pose estimation, thereby improving the efficiency of item classification, storage, and retrieval.

[0167] Building and Construction:

[0168] In the field of architecture and construction, this technology can be used to accurately identify and locate building materials and equipment, thereby improving the precision and efficiency of construction.

[0169] Archaeology and cultural relic protection:

[0170] In archaeology and cultural relic conservation, this technology can be used to accurately identify and estimate the position of cultural relics, thereby enabling better protection and restoration of cultural relics.

[0171] In summary, the FFT-based data augmentation method and optical flow-assisted multi-level alignment strategy proposed in this invention not only have significant advantages in 6D object pose estimation in underwater environments, but can also be widely applied to multiple application fields to improve the accuracy and robustness of object recognition and pose estimation in various application scenarios.

[0172] This invention also provides a storage medium for storing a computer program, which, when executed, performs at least the methods described above.

[0173] This invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein the processor executes the computer program by performing at least the method described above.

[0174] This invention also provides a processor that executes a computer program, at least performing the methods described above.

[0175] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); the magnetic surface memory can be a disk drive or magnetic tape drive. The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0176] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0177] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0178] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0179] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0180] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0181] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0182] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0183] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0184] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

Claims

1. A frequency-aware flow-assisted self-supervised method for underwater target pose estimation, characterized in that, Includes the following steps: S1. Acquire synthetic RGB images and real underwater captured RGB images; perform frequency domain analysis on the images using frequency domain transformation methods to extract amplitude and phase information, and execute frequency domain amplitude processing strategies to achieve data augmentation; Step S1 specifically includes: S11. Perform Fast Fourier Transform (FFT) on the collected synthetic RGB images and real underwater captured RGB images to convert the spatial domain images to frequency domain representation; S12. Separate the amplitude and phase information of the image from the FFT result, where the amplitude information reflects the brightness and contrast of the image, while the phase information contains the contour and edge features of the image; S13. Based on amplitude and phase information, perform amplitude blending operation on the image to combine the amplitude information of the synthesized image with the amplitude information of the real image in order to introduce style features of the target domain; S14. Implement an amplitude discarding strategy, randomly selecting a certain proportion of amplitude information for discarding; S15. Perform linear interpolation on the amplitude information. Use the amplitude information of two images randomly sampled from the synthetic dataset and the real-world dataset to generate a new amplitude spectrum by weighted averaging, so as to introduce style prior. S16. Based on a specific discard threshold, determine which amplitude information is retained and which is discarded in order to control the degree of amplitude spectrum discard; S17. Combining the transformed amplitude information and the original phase information, the image is reconstructed through inverse Fourier transform to generate an enhanced image for training, which integrates the style features of the synthetic image and the contour information of the real image. Repeat the above steps to generate an enhanced image set for subsequent network pre-training and self-supervised learning; S2. Using the synthetic image augmented by the data in step S1, pre-train a preliminary pose estimation network and a flow estimation network to obtain the initial pose and initial flow field of the image; use the preliminary pose estimation network to perform preliminary pose estimation on the real underwater image to obtain the initial pose prediction. S3. Using the given 3D model of the target object and the initial pose predicted in step S2, render multiple synthetic images that vary around the initial pose; S4. Based on the preliminary pose obtained in step S2 and the synthetic image generated in step S3, iterative optimization of pose and optical flow estimation is performed. This includes establishing a dense 2D-2D correspondence between the real and synthetic images and using optical flow to assist in image-level alignment; performing feature-level alignment in the high-dimensional latent space to extract and integrate high-level semantic information; and combining image-level and feature-level alignment with shape-constrained optical flow to perform self-supervised learning of the network.

2. The self-supervised method for underwater target pose estimation as described in claim 1, characterized in that, In step S1, the frequency domain amplitude processing strategy includes mixing and discarding amplitudes, or replacing amplitudes. The mixing and discarding of amplitudes includes: introducing style priors of the target domain by mixing phase information of the synthesized image with amplitude information of the real image, and randomly discarding amplitude information to enable the network to learn domain-invariant features.

3. The self-supervised method for underwater target pose estimation as described in claim 1 or 2, characterized in that, In step S2, based on the initial pose estimation network and flow estimation network, the teacher network and student network for the self-supervised learning stage are initialized. The teacher network is initialized using the parameters of the pre-trained pose estimation network and flow estimation network, and its network structure or parameters are fine-tuned to generate pseudo-labels to guide the training of the student network. The student network is initialized based on the parameters of the pre-trained pose estimation network and flow estimation network, and will be iteratively optimized independently of the teacher network during the self-supervised learning process. In step S4, during the iteration process, the teacher network and the student network work together. The teacher network generates pseudo-labels to assist the training of the student network. The pseudo-labels serve as self-supervised signals, guiding the training direction of the student network and providing auxiliary and consistency feedback.

4. The self-supervised method for underwater target pose estimation as described in any one of claims 1 to 2, characterized in that, Step S2 specifically includes: S21. Use the synthetic image obtained after data augmentation to initialize the training process of the pose estimation network and the flow estimation network; S22. Use the initial pose estimation network to predict the pose of the real underwater image to obtain the initial pose; S23. Calculate the corresponding ground flow based on the known ground pose; then, calculate the predicted flow based on the predicted initial pose. S24. Using supervised learning, combining real-world ground flow and predicted flow, calculate flow loss; S25. Using supervised learning, combine the real pose and the predicted pose to calculate the pose loss; S26. Determine the weight ratio of flow loss and pose loss; S27. Based on the determined weight ratio, combine the flow loss and pose loss to form the total supervision loss; S28. Use the total supervised loss to guide the training of the network and update the network parameters through the backpropagation algorithm; S29. Repeat steps S22 to S28 for multiple rounds of iterative training until the network achieves the predetermined accuracy in predicting flow and object pose.

5. The self-supervised method for underwater target pose estimation as described in any one of claims 1 to 2, characterized in that, Step S4 specifically includes: S41. Establish a dense pixel-level correspondence between the real image and each synthetic image to provide a basis for pose fine-tuning; S42. The network learns to predict shape constraint flow, that is, based on the correspondence, it predicts the motion of pixels under continuous pose iteration; S43. Combine the predicted shape constraint flow and camera parameters to iteratively update the pose estimation and gradually refine the pose prediction; S44. 2D-3D Correspondence Calculation: In each iteration, calculate the correspondence between visible object points in the input image and 3D model points; S45. Using the calculated 2D-3D correspondence, generate a pose guidance flow to guide further iterative optimization of the pose; S46. Decouple the rotation and translation components of the pose and estimate them separately; S47. A continuous rotation representation method is adopted to facilitate direct regression of pose parameters by the neural network; Repeat steps S42 to S47 for multiple iterations until the pose estimation reaches the predetermined accuracy.

6. The self-supervised method for underwater target pose estimation as described in any one of claims 1 to 2, characterized in that, In step S4, the image-level alignment specifically includes the following operations: The flow field prediction of the student network is obtained by using the student network to predict the flow field between pixels in real and synthetic images. The predicted flow field based on the student network is back-distorted on the real image and compared with the real image after pseudo-flow field transformation generated by the teacher network. Photometric consistency is used as a metric to calculate the back optical flow consistency loss. The consistency of the forward flow predicted by the student network on the warped rendering mask is evaluated by comparing the forward warping transformation results of the predicted flow field and the pseudo flow field and calculating the forward optical flow consistency loss. The teacher network is used to evaluate the consistency of the predicted flow between the synthetic and real images, and reliable pseudo-labels are screened out by standard deviation to supervise the flow prediction of the student network. The weighted sum of the backward optical flow consistency loss, the forward optical flow consistency loss, and the optical flow self-supervision loss yields the total loss for image-level alignment. Based on the image-level alignment loss, the parameters of the student network are updated through the backpropagation algorithm to optimize the accuracy of flow field prediction; Repeat the above steps for multiple iterations until the flow field prediction reaches the predetermined accuracy.

7. The self-supervised method for underwater target pose estimation as described in any one of claims 1 to 2, characterized in that, In step S4, the feature-level alignment specifically includes the following operations: A feature encoder using a pose network extracts feature representations from real underwater images and synthetic images, respectively. Based on the flow field predicted by the student network, the features of the real image are distorted to generate a distorted feature representation, taking into account the influence of environmental changes in the process. The difference between the distorted feature representation and the features of the synthesized image is calculated, and the feature-level loss is defined using a metric that can reflect feature similarity and adapt to environmental changes. Feature-level loss is used to guide the training of the pose network, and the network parameters are updated through the backpropagation algorithm; Repeat the above steps for multiple iterations until the alignment of the feature representations reaches the predetermined accuracy.

8. The self-supervised method for underwater target pose estimation as described in any one of claims 1 to 2, characterized in that, In step S4, pose loss and total self-supervised loss are introduced during the self-supervised learning process, including: The 3D point set uniformly sampled from the surface of the 3D model is transformed using pseudo-labeled pose and predicted pose to obtain the corresponding 2D point positions. Based on the transformed 2D point positions, the pose loss is calculated, which reflects the difference between the pseudo-labeled pose and the predicted pose. The L1 norm is used to measure the accuracy of point matching. The pose loss is used to guide the optimization of pose parameters, and the point matching loss is minimized by adjusting the pose parameters. Construct a total self-supervised loss, which is a weighted sum of pose loss, image-level loss, and feature-level loss, where associated weights are set for image-level loss and feature-level loss; Based on the calculated total self-supervised loss, the network parameters are updated using the backpropagation algorithm. Repeat the training process for multiple rounds until the network achieves the predetermined accuracy or the maximum number of iterations for the pose estimation task.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the self-supervised method for underwater target pose estimation as described in any one of claims 1-8.

Images