A single-point supervised panoramic target detection method based on spherical geometry modeling, a computer readable storage medium and a computer device

Abstract

The present application relates to panoramic target detection, weak supervision learning and spherical vision modeling technical field, specifically relates to a kind of single-point supervision panoramic target detection method based on spherical geometry modeling, computer readable storage medium and computer device, wherein method includes: S1, initial spherical boundary field of view candidate proposal generation;S2, scale-rotation enhancement is carried out on the basis of S1, multi-scale feature extraction, spherical region of interest feature extraction, multiple-instance learning head score, scale-rotation score consistency constraint, Top-k spherical weighted processing, obtain rough spherical boundary field of view;S3, on the basis of S2, spherical jitter, multi-scale positive candidate proposal generation, negative candidate proposal generation and iterative optimization processing are carried out, obtain pseudo spherical boundary field of view label training panoramic target detector for detection.The present application solves the problem that existing plane point supervision target detection method cannot adapt to equidistant cylindrical projection spherical geometry.

Description

Technical Field

[0001] This invention relates to the fields of panoramic object detection, weakly supervised learning, and spherical visual modeling. Specifically, it relates to a single-point supervised panoramic object detection method based on spherical geometric modeling, a computer-readable storage medium, and a computer device. Background Technology

[0002] With the development of panoramic imaging technology, panoramic images have been applied in fields such as virtual reality, autonomous driving, intelligent robot navigation, and street view mapping. Compared with ordinary images, panoramic images can provide more complete environmental information, thus the research demand for panoramic object detection is constantly increasing.

[0003] Current panoramic object detection methods primarily use isometric cylindrical projection images as input and employ spherical bounding field of view (BFoV) to describe the target. However, existing panoramic object detection datasets are limited, and isometric cylindrical projection images exhibit significant geometric distortion. To accurately annotate targets, annotators need to use BFoV and continuously rotate the viewing angle to confirm the target's extent, resulting in high annotation costs and difficulties.

[0004] Existing single-point supervised object detection techniques can generate planar proposal boxes of different scales and proportions centered on labeled points. They then use a two-stream multi-instance learning classifier with both classification and instance branches to score the proposal boxes, select several high-scoring proposal boxes for weighted averaging, and generate improved proposal boxes multiple times around the initial pseudo-boxes. The prior art, with publication number CN115100501A, also discloses a precise object detection method based on single-point supervision, which trains a second classifier using negative example boxes with an intersection-union ratio (IU) below a preset threshold. Therefore, single-point annotation, two-stream multi-instance learning, Top-q candidate box weighting, and coarse-to-fine iterative fine-tuning are already existing technical approaches in planar point supervised object detection.

[0005] In the aforementioned existing technologies, candidate boxes are represented by planar center coordinates, width, and height. Candidate region features are typically obtained using the planar RoIAlign algorithm. Candidate box aggregation employs a Euclidean weighted average of center coordinates, width, and height, while the fine-tuning stage adjusts the center position through planar coordinate offset. Applying this directly to isometric cylindrical projection panoramic images, even after converting the planar pseudo-boxes to BFoV, does not alter the planar geometric relationships underlying candidate region feature extraction, candidate center aggregation, and candidate position fine-tuning.

[0006] Equidistant cylindrical projection maps spherical longitude and latitude to planar coordinates. The same spherical target exhibits different degrees of planar stretching at different latitudes, and the left and right boundaries of the image correspond to continuous positions on the sphere. Therefore, using a plane circumscribed rectangle for RoIAlign will infiltrate background regions unrelated to the spherical target; directly averaging latitude and longitude or plane center coordinates will introduce center deviations in cross-boundary and mid-to-high latitude regions; and moving candidate centers according to a fixed number of pixels or direct latitude and longitude increments will cause inconsistencies in the actual east-west spherical displacement due to meridian convergence.

[0007] Furthermore, conventional data augmentation only changes the input image and does not explicitly constrain the consistency of the scores of the same spherical candidate proposals before and after scale changes and 3D spherical rotation. Therefore, it is difficult to guarantee that the ranking of candidate proposals remains stable under different spherical positions and observation scales. Summary of the Invention

[0008] In view of this, the present invention proposes a single-point supervised panoramic target detection method, a computer-readable storage medium, and a computer device based on spherical geometry modeling. It aims to solve the problem that existing planar point supervised target detection methods cannot adapt to equidistant cylindrical projection spherical geometry, and provides a single-point supervised panoramic target detection scheme that enables candidate proposal generation, region of interest feature extraction, Top-k center aggregation, and fine position optimization to be executed collaboratively under spherical geometry constraints.

[0009] This invention proposes a single-point supervised panoramic target detection method based on spherical geometry modeling, comprising: S1. Initial spherical boundary field of view candidate proposal generation: Obtain the isometric cylindrical projection panoramic image of the original view and the single-point annotation of the target in the isometric cylindrical projection panoramic image, convert the single-point annotation into a spherical position on a unit sphere, and generate an initial spherical boundary field of view candidate proposal with the spherical position as the candidate proposal center. S2. Coarse-grained field-of-view proposal generation: Based on S1, scale-rotation enhancement is performed on the original view and its initial spherical boundary field-of-view candidate proposals to obtain a scale-rotation enhanced view and enhanced candidate proposals corresponding one-to-one with the initial spherical boundary field-of-view candidate proposals; multi-scale feature extraction is performed on the original view and the scale-rotation enhanced view respectively; for the initial spherical boundary field-of-view candidate proposals and the enhanced candidate proposals, spherical region of interest features are extracted based on local tangent plane dense sampling, spherical backprojection, and feature pyramid hierarchical allocation; the spherical region of interest features are input into a shared multi-instance learning head for scoring, and scale-rotation score consistency constraints are applied to the scoring results of the initial spherical boundary field-of-view candidate proposals and the enhanced candidate proposals; Top-k spherical weighting is performed based on the final score of the initial spherical boundary field-of-view candidate proposals in the original view to obtain a coarse spherical boundary field of view; S3, Fine-grained field of view proposal optimization: Based on S2, spherical jitter, multi-scale positive candidate proposal generation, negative candidate proposal generation, and iterative optimization are performed according to the coarse spherical boundary field of view to obtain pseudo-spherical boundary field of view labels; the panoramic target detector is trained using the pseudo-spherical boundary field of view labels; the isometric cylindrical projection panoramic image to be detected is input into the panoramic target detector, and the target category and the corresponding spherical boundary field of view are output; The total loss function for training the pseudo-spherical boundary field of view generation network is as follows: ; in, and The coarse-grained multi-instance learning loss is used for the original view and the scale-rotation augmented view, respectively. For the fine-grained stage of multi-instance learning loss, For negative proposal losses, For scale-rotation fractional consistency loss, and This is the balance coefficient.

[0010] Preferably, the generation of initial spherical boundary field of view candidate proposals in S1 specifically includes: The single-point annotation is converted into a spherical position on a unit sphere, and the spherical position corresponding to the single-point annotation is used as the candidate proposal center. A local tangent plane is established at the candidate proposal center, and a minimum proposal scale is set with both the horizontal and vertical field of view angles of 6°. Multi-scale candidate regions are generated on the local tangent plane, and the multi-scale candidate regions are back-projected onto the unit sphere to obtain the spherical boundary field of view candidate proposal set of the original view. The spherical boundary field of view candidate proposal set This is the initial set of candidate proposals for the spherical boundary field of view.

[0011] Preferably, the scale-rotation enhancement and multi-scale feature extraction processing in S2 specifically includes: The original view is subjected to a planar scale transformation, and the same three-dimensional spherical rotation is applied to the original view and its initial spherical boundary field of view candidate proposals to obtain a scale-rotation enhanced view and a set of enhanced candidate proposals that correspond one-to-one with the initial spherical boundary field of view candidate proposals. The original view and the scale-rotation augmented view are respectively input into the backbone network and the feature pyramid network with shared weights to obtain the multi-scale features of the original view and the multi-scale features of the scale-rotation augmented view.

[0012] Preferably, the spherical interest feature extraction process in S2 specifically includes: Based on the spherical geometric parameters of each spherical boundary field of view candidate proposal, a rotatable dense sampling mesh is established on the corresponding local tangent plane; a rotated spherical boundary field of view is then used. Define a local sampling region, where, and Indicates the center of the sphere. and Indicates the horizontal and vertical field of view. Indicates the rotation angle within the tangent plane; when When the rotating spherical boundary field of view degenerates into the spherical boundary field of view, the rotating spherical boundary field of view is only used to define the local sampling region, and the final generated pseudo label is the spherical boundary field of view; At the center of the field of view of the boundary of the rotating spherical surface Construct a central unit vector and tangential basis vectors along local parallels of latitude and longitude. and Let the horizontal half-range of the tangent plane Vertical half range Set the sampling grid resolution. , For sampling density, and ;No. The coordinates of the tangent plane of each sampling point are: ; The first The coordinates of the tangent plane at each sampling point rotate around the center of the candidate proposals in the spherical boundary field of view. : ; Back-project the rotated sampling points onto the unit sphere: ; According to spherical coordinates Obtain the continuous pixel coordinates on the isometric cylindrical projection feature map of the corresponding view; the geometric scale of the spherical boundary field of view candidate proposals. and feature pyramid hierarchy They are respectively: ; ; in, The width of the isometric cylindrical projection panoramic image of the current processing view at the input scale; For reference scale; As the baseline feature level; from the first The layered feature pyramid layer uses bilinear interpolation to extract densely sampled features and aggregates these features into... Features of the region of interest on a sphere.

[0013] Preferably, the scale-rotation fraction consistency constraint in S2 specifically includes: The initial spherical boundary field of view candidate proposals of the original view and their corresponding augmentation candidate proposals in the scale-rotation augmented view are input into a shared multi-instance learning head to obtain instance scores and class scores; for any view, the first... One candidate proposal, multi-instance learning head output instance score and category scores The final score for the candidate proposals is : ; in, Represents element-wise product; The original view candidate proposal group and its scale-rotation augmented view candidate proposal group corresponding to the same single-point annotation are regarded as a corresponding candidate proposal group. Within each corresponding candidate proposal group, the cosine distance between the initial spherical boundary field of view candidate proposal in the original view and the augmented candidate proposal in the scale-rotation augmented view is calculated as follows: ; in, Indicates a single-point annotation. Indicates the candidate proposal group, This represents the candidate proposals in the original view; Indicates scale - rotation enhancement view and Corresponding enhancement candidate proposals; Represents a fraction type and includes instance fraction types. and category score type The cosine distance is minimized using a smooth L1 loss: ; in, Indicates smoothed L1 loss, and This indicates the balancing weight.

[0014] Preferably, the Top-k spherical weighting in S2 specifically includes: Based on the final score of the initial spherical boundary field of view candidate proposals in the original view Select the top-k candidate proposals; normalize the scores of the top-k candidate proposals into weights. ,in Indicates the first The normalized weights corresponding to the Top-k candidate proposals; the spherical centers of each Top-k candidate proposal are converted into three-dimensional unit vectors. ,in Indicates the first The unit vector representation of the spherical center of each of the top-k candidate proposals in three-dimensional Euclidean space. Represents the unit sphere in three-dimensional Euclidean space; Using the weighted sum of normalized three-dimensional unit vectors as the initial center :

[0015] ; Iterative updates are performed using exponential and logarithmic mappings: ; in, Indicated by Spherical exponential mapping with reference to the base point; Indicated by Mapping of the number of spherical surfaces with reference to the base point; Indicates the step size; when the spherical distance between two consecutive iterations is less than the threshold. Or reach the maximum number of iterations Stop at time, and obtain the weighted spherical center. ; The horizontal and vertical field of view of the candidate proposals are weighted according to the following formula: ; in, and They represent the first Horizontal and vertical field of view of the Top-k candidate proposals and These represent the converged horizontal and vertical field of view, respectively; centered on the weighted sphere. The converted latitude and longitude centers and and Together they form a rough spherical boundary field of view.

[0016] Preferably, S3 specifically includes: Using the coarse spherical boundary field of view as the initial value, the angular displacement is determined according to the horizontal and vertical field of view angles of the coarse spherical boundary field of view. The spherical surface is shaken in four directions along the lateral great circle direction and the longitudinal meridian direction. Multi-scale positive candidate proposals are generated at each shaking center, and negative candidate proposals with a spherical intersection-union ratio of less than 0.3 are generated. Given the center of the field of view of the spherical boundary Horizontal field of view Vertical field of view and the ratio of shaking ,make: ; When vibrating along the horizontal large circle: ; When vibrating along the longitudinal meridian: ; in, and This indicates the latitude and longitude of the center of the sphere after shaking. When the center of the sphere after shaking crosses the poles, boundary processing is performed. Positive and negative candidate proposals are processed through a shared spherical region of interest feature extraction module and a multi-instance learning head. The prediction scores of positive candidate proposals are normalized and used as spherical weights to update the spherical boundary field of view center and field of view angle. The responses of negative candidate proposals are suppressed by negative proposal loss. The updated spherical boundary field of view generates spherical jitter and multi-scale candidate proposals again. The candidate proposal generation, scoring and spherical weighting process is repeated to obtain the final pseudo-spherical boundary field of view label. The panoramic target detector is trained using the pseudo-spherical boundary field of view label.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention does not approximate the spherical candidate region as the circumscribed rectangle of the equidistant cylindrical projection plane, but directly constructs a dense sampling grid on the local tangent plane and back-projects it onto the unit sphere, extracting features on the continuous equidistant cylindrical projection coordinates; based on the field scale of the tangent plane, the feature pyramid hierarchy is allocated, making the candidate region features more stable to latitude changes, and eliminating the need for mask generation, morphological screening and extraction of the circumscribed box of the maximum connected region.

[0018] (2) The present invention maintains the stability of candidate proposal scores before and after scale changes and 3D spherical rotation by constraining the consistency of instance scores and class scores of corresponding candidate proposals in the original view and scale-rotation enhanced view; the enhanced view only provides training supervision and does not affect the Top-k aggregation and subsequent fine optimization of the original view.

[0019] (3) The present invention places the candidate center on the unit spherical manifold and calculates the Top-k weighted center through logarithmic mapping, tangent space weighting and exponential mapping, avoiding cross-boundary deviation and mid-to-high latitude geometric error caused by directly averaging latitude and longitude or planar coordinates; the field of view is independently weighted in the local tangent plane parameter space.

[0020] (4) The present invention determines the angular displacement based on the current BFoV field of view, performs four-directional spherical jitter along the transverse great circle and longitudinal meridian, and iteratively combines spherical jitter, candidate score and spherical weighting to avoid excessive east-west displacement in high latitude regions due to direct latitude and longitude offset.

[0021] (5) This invention enables the initial candidate generation, feature extraction, candidate center aggregation and fine position optimization to be performed under a unified spherical geometric relationship. While reducing the cost of manual annotation of complete BFoV, it improves the positioning accuracy of pseudo BFoV labels, especially for mid-to-high latitude targets. The generated pseudo BFoV labels can be used to train panoramic target detectors with different structures, and single-point annotation is not required in the deployment and inference stage.

[0022] On the other hand, this application provides a computer device including a processor and a memory. The memory stores program instructions, parameters of a trained pseudo-spherical boundary field of view generation network, and parameters of a panoramic target detector. When the program instructions are executed by the processor, the single-point supervised panoramic target detection method based on spherical geometry modeling is executed.

[0023] Preferably, in the pseudo-spherical boundary field of view label generation stage, the processor loads the pseudo-spherical boundary field of view generation network parameters and generates pseudo-spherical boundary field of view labels based on the equidistant cylindrical projection panoramic image and single-point annotations; in the panoramic target detection deployment stage, the processor loads the panoramic target detector parameters, takes the equidistant cylindrical projection panoramic image to be detected as input, and outputs the target category and the corresponding spherical boundary field of view.

[0024] It is understood that the computer device provided by this invention has the same beneficial effects as a single-point supervised panoramic target detection method based on spherical geometry modeling, and will not be described in detail here.

[0025] In another aspect, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned single-point supervised panoramic target detection method based on spherical geometry modeling.

[0026] It is understood that the computer-readable storage medium provided by this invention has the same beneficial effects as a single-point supervised panoramic target detection method based on spherical geometry modeling, and will not be elaborated here. Attached Figure Description

[0027] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart illustrating a single-point supervised panoramic target detection method based on spherical geometry modeling, provided in an embodiment of the present invention. Detailed Implementation

[0028] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0029] See Figure 1 This embodiment provides a single-point supervised panoramic target detection method based on spherical geometry modeling, including: S1. Initial Spherical Boundary Field of View Candidate Proposal Generation: Obtain the isometric cylindrical projection panoramic image (ERP image) of the original view and the single-point annotation of the target in the isometric cylindrical projection panoramic image. Convert the single-point annotation into a spherical position on a unit sphere and generate the initial spherical boundary field of view candidate proposal with the spherical position as the candidate proposal center.

[0030] In this embodiment, the input original image is an isometric cylindrical projection panoramic image, and each target in the isometric cylindrical projection panoramic image is provided with only a single-point annotation. The single-point annotation is used to represent the approximate spherical center position of the target and serves as input for generating candidate proposals.

[0031] Preferably, the generation of initial spherical boundary field of view candidate proposals in S1 specifically includes: Single-point annotations are converted into spherical positions on a unit sphere, and these spherical positions are used as candidate proposal centers. A local tangent plane is established at the candidate proposal center, with both the horizontal and vertical field of view angles set to 6° as the minimum proposal scale. Multi-scale candidate regions are generated on the local tangent plane, and these regions are back-projected onto the unit sphere to obtain the spherical boundary field of view candidate proposal set of the original view. Set of candidate proposals for spherical boundary field of view This is the initial set of candidate proposals for the spherical boundary field of view.

[0032] S2. Coarse-grained field-of-view proposal generation: Based on S1, scale-rotation enhancement is performed on the original view and its initial spherical boundary field-of-view candidate proposals to obtain a scale-rotation enhanced view and enhancement candidate proposals that correspond one-to-one with the initial spherical boundary field-of-view candidate proposals; multi-scale feature extraction is performed on the original view and the scale-rotation enhanced view respectively; for the initial spherical boundary field-of-view candidate proposals and enhancement candidate proposals, spherical region of interest features are extracted based on local tangent plane dense sampling, spherical backprojection, and feature pyramid hierarchical allocation; the spherical region of interest features are input into a shared multi-instance learning head for scoring, and scale-rotation score consistency constraints are applied to the scoring results of the initial spherical boundary field-of-view candidate proposals and enhancement candidate proposals; Top-k spherical weighting is performed based on the final score of the initial spherical boundary field-of-view candidate proposals in the original view to obtain a coarse spherical boundary field of view.

[0033] Preferably, the scale-rotation enhancement and multi-scale feature extraction processing in S2 specifically includes: A planar scale transformation is performed on the original view, and the same three-dimensional spherical rotation is applied to the original view and its initial spherical boundary field of view candidate proposals to obtain a scale-rotation enhanced view and a set of enhanced candidate proposals that correspond one-to-one with the initial spherical boundary field of view candidate proposals. The original view and the scale-rotation augmented view are respectively input into the backbone network and the feature pyramid network with shared weights to obtain the multi-scale features of the original view and the multi-scale features of the scale-rotation augmented view.

[0034] The original view and the scale-rotation augmented view are respectively input into a backbone network (e.g., ResNet) and a feature pyramid network (FPN) with shared weights to extract multi-scale features from layers P2, P3, P4 and P5, thus obtaining multi-scale features.

[0035] Specifically, the original view and the scale-rotation augmented view use the same backbone network, feature pyramid network, spherical region of interest feature extraction module, and multi-instance learning head. However, the scale-rotation augmented view only participates in the coarse-grained training stage, does not perform Top-k aggregation, and does not enter the fine field of view proposal optimization stage.

[0036] Preferably, the spherical interest feature extraction process in S2 specifically includes: Based on the spherical geometric parameters of each spherical boundary field of view candidate proposal, a rotatable dense sampling mesh is established on the corresponding local tangent plane; Using a rotating spherical boundary field of view Define a local sampling region, where, and Indicates the center of the sphere. and Indicates the horizontal and vertical field of view. Indicates the rotation angle within the tangent plane; when When the rotating spherical boundary field of view degenerates into the spherical boundary field of view, the rotating spherical boundary field of view is only used to define the local sampling region, and the final generated pseudo label is the spherical boundary field of view; At the center of the field of view of the boundary of the rotating spherical surface Construct a central unit vector and tangential basis vectors along local parallels of latitude and longitude. and Let the horizontal half-range of the tangent plane Vertical half range ; Set sampling grid resolution , The sampling density corresponding to each final pooling unit, and ;No. The coordinates of the tangent plane of each sampling point are: ; The first The coordinates of the tangent plane at each sampling point rotate around the center of the candidate proposals in the spherical boundary field of view. : ; Back-project the rotated sampling points onto the unit sphere: ; According to spherical coordinates Obtain the continuous pixel coordinates on the isometric cylindrical projection feature map of the corresponding view; the geometric scale of the spherical boundary field of view candidate proposals. and feature pyramid hierarchy They are respectively: ; ; in, The width of the isometric cylindrical projection panoramic image of the current processing view at the input scale; For reference scale; As the baseline feature level; from the first The layered feature pyramid layer uses bilinear interpolation to extract densely sampled features and aggregates these features into... Features of the region of interest on a sphere.

[0037] In this embodiment, when extracting spherical region of interest features based on local tangent plane dense sampling, spherical backprojection, and feature pyramid hierarchical allocation for initial spherical boundary field of view candidate proposals and enhanced candidate proposals, it is not necessary to generate candidate region masks, perform morphological filtering, or extract plane bounding rectangles on the equidistant cylindrical projection image. This embodiment directly generates the sampling mesh based on the spherical parameters of BFoV or RBFoV, without generating equidistant cylindrical projection region masks, or performing morphological filtering, maximum connected region extraction, or bounding rectangle RoIAlign. This ensures that the sampling region corresponds to the spherical field of view of the target, avoiding excessive background inclusion in the plane bounding rectangle in high-latitude regions. RBFoV (Rotated Spherical Boundary Field of View) is only used to define rotatable local sampling regions. The final generated and saved pseudo BFoV (Spherical Boundary Field of View) labels are still standard BFoVs without in-tangent plane rotation angles. Using RBFoV sampling ensures that the local mesh orientation is consistent with the candidate region orientation, while not changing the BFoV label format used by the downstream panoramic target detector.

[0038] Furthermore, in this embodiment, the feature pyramid level of the candidate proposal is determined by the horizontal and vertical half-ranges of the tangent plane, rather than by the pixel width and height of the candidate region in the equidistant cylindrical projection plane. Since the scale of the tangent plane is defined by the horizontal and vertical field of view, the same spherical field of view can be assigned to a consistent scale level at different latitudes; after dense sampling, it is aggregated into a fixed 7×7 feature, which reduces aliasing caused by sparse sampling while keeping the input size of the subsequent multi-instance learning head unchanged.

[0039] Specifically, in this embodiment, the dense sampling resolution Where s represents the sampling density corresponding to each final pooling unit. When s increases from 1 to 2, aliasing can be reduced and fine-grained texture can be preserved by denser tangent sampling under a fixed 7×7 output size; when s continues to increase to 3, the number of sampling points increases quadratically. Considering both positioning accuracy and computational overhead, this embodiment uses s=2.

[0040] Preferably, the scale-rotation fraction consistency constraint in S2 specifically includes: The initial spherical boundary field of view candidate proposals of the original view and their corresponding augmentation candidate proposals in the scale-rotation augmented view are input into a shared multi-instance learning (MIL) head to obtain instance scores and class scores; for any view, the first... One candidate proposal, multi-instance learning head output instance score and category scores The final score for the candidate proposals is : ; in, Represents element-wise product; The original view candidate proposal group and its scale-rotation augmented view candidate proposal group corresponding to the same single-point annotation are regarded as a corresponding candidate proposal group. Within each corresponding candidate proposal group, the cosine distance between the initial spherical boundary field of view candidate proposal in the original view and the augmented candidate proposal in the scale-rotation augmented view is calculated as follows: ; in, Indicates a single-point annotation. Indicates the candidate proposal group, This represents the candidate proposals in the original view; Indicates scale - rotation enhancement view and Corresponding enhancement candidate proposals; Represents a fraction type and includes instance fraction types. and category score type The cosine distance is minimized using a smooth L1 loss: ; in, Indicates smoothed L1 loss, and This indicates the balancing weight.

[0041] In this embodiment, the scale-rotation score consistency constraint is applied to the instance score distribution and the class score distribution, respectively, rather than only constraining the final detection results of the two views. By minimizing the cosine distance of the corresponding candidate proposal groups, the network can maintain the stability of the candidate proposal ranking and class judgment after scale changes and 3D spherical rotation, enabling the enhanced view to provide transformation-aware regularized supervision.

[0042] Furthermore, when constructing the scale-rotation enhanced view, the same 3D spherical rotation is performed on the original view and its initial spherical boundary field of view candidate proposals, and the image is scaled to achieve the desired spherical boundary field of view candidate proposal set. With the enhancement of the candidate proposal set The candidate proposals in the view are kept in the order they were generated. Both views are extracted and scored using a shared weighted feature extraction and scoring module, so the score differences mainly reflect the response changes caused by scale and rotation transformations.

[0043] The instance branch is used to distinguish candidate regions within the same set of candidate proposals, while the category branch is used to determine the target category. Constraining the score distributions of the two branches separately avoids the error cancellation that occurs when only constraining the final score after the Hadamard product of the two branches. The scale-rotation augmented view is only used to calculate the coarse-grained multi-instance learning loss and the scale-rotation score consistency loss; Top-k candidate selection, spherical weighting, and fine-grained field-of-view proposal optimization are all based on the original view.

[0044] The scale-rotation score consistency constraint in this embodiment differs from simply applying random augmentation to the input: it utilizes the explicit correspondence between the same candidate proposal before and after the transformation to directly supervise the instance score and category score on which the candidate proposal ranking is based, thereby enhancing the stability of the candidate score with respect to scale and spherical rotation.

[0045] Preferably, the Top-k spherical weighting in S2 specifically includes: Based on the final score of the initial spherical boundary field of view candidate proposals in the original view Select the top-k candidate proposals; normalize the scores of the top-k candidate proposals into weights. ,in Indicates the first The normalized weights corresponding to the Top-k candidate proposals; the spherical centers of each Top-k candidate proposal are converted into three-dimensional unit vectors. ,in Indicates the first The unit vector representation of the spherical center of each of the top-k candidate proposals in three-dimensional Euclidean space. Represents the unit sphere in three-dimensional Euclidean space; Using the weighted sum of normalized three-dimensional unit vectors as the initial center :

[0046]

[0047] Iterative updates are performed using exponential and logarithmic mappings: ; in, Indicated by Spherical exponential mapping with reference to the base point; Indicated by Mapping of the number of spherical surfaces with reference to the base point; Indicates the step size; when the spherical distance between two consecutive iterations is less than the threshold. Or reach the maximum number of iterations Stop at time, and obtain the weighted spherical center. ; The horizontal and vertical field of view of the candidate proposals are weighted according to the following formula: ; in, and They represent the first Horizontal and vertical field of view of the Top-k candidate proposals and These represent the converged horizontal and vertical field of view, respectively; centered on the weighted sphere. The converted latitude and longitude centers and and Together they form a rough spherical boundary field of view.

[0048] In a preferred embodiment, the maximum number of iterations T is set to 30, and the convergence threshold ε is set to 10. -4 This setting is only used to limit the stopping condition of the spherical center iteration; the horizontal and vertical field of view angles are still weighted as scalars on the local tangent plane.

[0049] In this embodiment, the spherical weighting differs from the arithmetic mean of longitude, latitude, and planar center coordinates. Each candidate center is first transformed to the tangent space of the current iteration center through a logarithmic mapping. Within the tangent space, weights are applied based on the candidate proposal scores, and then an exponential mapping is used to return the result to the unit sphere. If the arithmetic mean of longitude and latitude is directly applied, candidate centers located on either side of the equidistant cylindrical projection boundaries but adjacent on the sphere may obtain average positions far from the true target. If only a single Euclidean weighting and normalization is performed on the three-dimensional unit vector, the result is an external mean, which cannot explicitly minimize the spherical geodesic distance. This embodiment uses the normalized three-dimensional weighted sum as the initial value for iteration, and then obtains the weighted Karcher center through logarithmic and exponential mappings on the unit sphere. Thus, the center update distance is consistent with the spherical geodesic distance, avoiding the average center shift caused by the longitude cycle boundary and the convergence of mid-to-high latitude meridians.

[0050] Furthermore, in each iteration, the logarithmic map transforms each candidate center into a geodesic displacement vector in the tangent space of the current center. This vector is then summed according to the candidate score normalized weights, multiplied by the step size ρ, and finally returned to the unit sphere by the exponential map. The iteration stops when the spherical distance between adjacent iteration centers is less than ε. This embodiment uses... , The average number of actual iterations was 8.51, and the convergence rate was 97.9%.

[0051] The center of the sphere and the field of view use different weighting spaces: the center, belonging to a unit spherical manifold, uses geodesic weighting; the horizontal and vertical field of view are scalars on local tangent planes, using ordinary weighted averages. This avoids uniformly applying Euclidean averages to parameters with different geometric properties.

[0052] In this embodiment S2, the spherical Region of Interest (RoI) feature extraction first extracts features matching the spherical region for each BFoV candidate proposal. Scale-rotation score consistency constraints ensure that the feature maintains a stable score before and after geometric transformation after passing through a shared multi-instance learning head. Top-k spherical weighting is then used to fuse candidate centers on a unit sphere based on the scores. After the coarse BFoV enters the fine stage, spherical jitter generates new candidate centers along the geodesic direction. The new candidates still use the same spherical RoI and multi-instance learning head for scoring and are updated using the same spherical weighting method. Therefore, the distinguishing features are not independently superimposed but rather act separately on the feature representation, score stability, center fusion, and location search of the candidate proposals, and are connected end-to-end in the iterative process.

[0053] S3, Fine Field of View Proposal Optimization: Based on S2, spherical jitter, multi-scale positive candidate proposal generation, negative candidate proposal generation, and iterative optimization are performed according to the coarse spherical boundary field of view to obtain pseudo-spherical boundary field of view labels; the panoramic target detector is trained using the pseudo-spherical boundary field of view labels; the isometric cylindrical projection panoramic image to be detected is input into the panoramic target detector, and the target category and the corresponding spherical boundary field of view are output. In the coarse-grained field-of-view proposal generation stage, the original view and the scale-rotation enhanced view are supervised by the first multi-instance learning loss. In the fine-grained field-of-view proposal optimization stage, only the original view is used, with positive candidate proposals supervised by the second multi-instance learning loss and negative candidate proposals supervised by the negative proposal loss. The total loss function for training the pseudo-spherical boundary field-of-view generation network is as follows: ; in, and The coarse-grained multi-instance learning loss is used for the original view and the scale-rotation augmented view, respectively. For the fine-grained stage of multi-instance learning loss, For negative proposal losses, For scale-rotation fractional consistency loss, and This is the balance coefficient.

[0054] After training, a pseudo-spherical boundary field of view generation network is used to generate pseudo-BFoV labels based on single-point annotations, and these pseudo-BFoV labels are then used to train the panoramic object detector. During the inference phase, the panoramic object detector only requires isometric cylindrical projection panoramic images as input and does not need single-point annotations.

[0055] Specifically, the spherical dithering moves the candidate center along the spherical geodesic lines, rather than directly translating the latitude and longitude coordinates in the equidistant cylindrical projection plane.

[0056] Preferably, S3 specifically includes: Using a coarse spherical boundary field of view as the initial value, the angular displacement is determined based on the horizontal and vertical field of view angles of the coarse spherical boundary field of view. Four-directional spherical dithering is then performed along the lateral great circle direction and the longitudinal meridian direction of the sphere, generating multi-scale positive candidate proposals at each dithering center, and generating negative candidate proposals with a spherical intersection-union ratio (IUU) lower than 0.3. The specific formula is as follows: Given the center of the field of view of the spherical boundary Horizontal field of view α, vertical field of view β, and jitter ratio ,make , ; When vibrating along the horizontal large circle: ; When vibrating along the longitudinal meridian: ; Positive and negative candidate proposals are processed using shared spherical region of interest feature extraction and a multi-instance learning head. The prediction scores of positive candidate proposals are normalized and used as spherical weights to update the BFoV center and field of view. The responses of negative candidate proposals are suppressed by negative proposal loss. The updated BFoV generates spherical jitter and multi-scale candidate proposals again. The candidate proposal generation, scoring, and spherical weighting processes are repeated to obtain the final pseudo BFoV labels. The panoramic target detector is trained using the pseudo spherical boundary field of view labels.

[0057] Specifically, in this embodiment, negative candidate proposals are generated for each positive candidate proposal bag. When the spherical intersection-union ratio (SUIR) between a candidate proposal and its corresponding positive candidate proposal is less than 0.3, it is considered a negative candidate proposal. Negative candidate proposals provide background supervision and do not participate in spherical weighting.

[0058] Specifically, in this embodiment, the lateral jitter is achieved by using a great circle passing through the current center, instead of directly increasing the longitude while keeping the latitude constant. Since the spacing between meridians decreases with increasing latitude, a direct longitude offset would produce different spherical displacements at different latitudes; the great circle calculation transforms the jitter distance from angular displacement... Confirmed. The longitudinal shaking occurs along the meridian, and its angular displacement is determined by… Sure.

[0059] Angular displacement is proportional to the current horizontal or vertical field of view of the BFoV, thus the search range adaptively changes with the target field of view size. When the jitter center crosses the poles, spherical boundary processing is performed to ensure that the updated center still corresponds to a valid position on a unit sphere. Spherical jitter only changes the candidate center, subsequently generating multiple field of view scales at each center to simultaneously search for the center position and target range.

[0060] It is understood that the multi-instance learning dual-stream scoring, negative proposal supervision, and coarse-to-fine iteration in this embodiment can have similar forms to existing planar point supervision frameworks. However, this embodiment does not use planar rectangular boxes and their Euclidean coordinates as geometric computation objects. The initial proposal is a spherical BFoV, and the candidate region features are obtained by spherical back-projection sampling. The Top-k centers converge on the unit spherical manifold, and the refined candidate centers move along the geodesic direction of the spherical surface. The corresponding candidate proposals are subject to score consistency constraints before and after the scale-rotation transformation. The above-mentioned spherical technical features work together to solve the problems of latitudinal distortion, periodic boundaries, and meridian convergence in isometric cylindrical projection images.

[0061] Furthermore, the method in this embodiment generates multi-scale BFoV candidate proposals on a sphere based on single-point annotations. A coarse BFoV is obtained through spherical region of interest feature extraction, multi-instance learning scoring, scale-rotation score consistency constraints, and Top-k spherical weighting. Then, pseudo-BFoV labels are obtained through spherical jittering, multi-scale candidate proposal generation, negative proposal constraints, and iterative spherical weighting. These pseudo-BFoV labels are then used to train the panoramic target detector. Compared to complete BFoV annotation, single-point annotation reduces annotation costs. Local tangent plane sampling, spherical weighting, and spherical jittering follow spherical geometry, mitigating the impact of projection distortion in equidistant cylindrical projection images on candidate proposal localization. This addresses issues such as inconsistent features across candidate regions at different latitudes, the inability to directly average spherical candidate centers using Euclidean methods, the inability of fine-tuning planar coordinates to maintain equisional spherical angular distances, and instability in candidate scores before and after enhancement. This improves the localization quality of pseudo-BFoV labels in mid-to-high latitude regions.

[0062] Experimental setup, feature validation, and results: This embodiment was tested on a 360-Indoor panoramic dataset. The dataset contains 37 object categories, 2325 training images, and 1010 test images. The evaluation metrics are mAP and AP50, and the spherical intersection-union ratio is calculated using the Sph2Pob fast approximation method. During training and inference, the panoramic images are uniformly scaled to 1024×512. The network is implemented based on MMDetection, using a stochastic gradient descent optimizer and a 1× training plan. The initial learning rate is 0.02, which decays to 0.1 in rounds 8 and 11, respectively.

[0063] To simulate low-cost point annotation, pseudo-realistic single-point annotations were randomly sampled within a spherical angle range of 0.1 times the corresponding field of view, using the center of the real BFoV sphere as the mean of a Gaussian distribution. The generated pseudo-BFoV labels were used to train the panoramic object detector, which does not require point annotations during inference.

[0064] Under the same annotation budget, based on the annotation time estimates for 100 single-point annotations and 100 BFoV annotations, the number of BFoV annotations used in the fully supervised method is set to 1 / 7 of the number of point annotations. Sph-CenterNet trained using the pseudo-BFoV labels of this invention achieves 7.33% mAP and 18.51% AP50, higher than the 6.01% and 14.19% of Sph-CenterNet trained using limited manual BFoV annotations.

[0065] In the comparison of different point supervision methods, the pseudo BFoV generated in this embodiment achieved a mAP of 9.56% and an AP50 of 31.04%, which is higher than P2BNet-BFoV's 5.95% and 20.42%, and PointOBB-BFoV's 4.37% and 17.11%. The pseudo BFoV generated in this embodiment is closer to the real BFoV and has higher positioning accuracy and smaller position deviation in mid-to-high latitude regions.

[0066] To verify the effectiveness of this embodiment in distinguishing features from planar point supervision techniques, further comparisons were made on spherical region of interest feature extraction, feature quality at different latitudes, spherical weighting and spherical jitter, sampling density, spherical weighting iteration parameters, and scale-rotation fractional consistency loss. All experimental data described below are from the same 360-Indoor experimental setup.

[0067] Table 1 Comparison of different region of interest feature extraction strategies Box RoI 3.70 0.43 8.54 24.69 Sphere RoI 2.39 0.48 9.56 31.04 In Table 1, Box RoI first generates a BFoV region mask on an equidistant cylindrical projection plane, filters and extracts the bounding rectangle of the largest connected region, and then executes RoIAlign; Sphere RoI directly constructs a tangent plane sampling mesh from BFoV geometry. Sphere RoI improves mIoU from 0.43 to 0.48, mAP from 8.54% to 9.56%, and AP50 from 24.69% to 31.04%, while reducing the training time per round from 3.70 hours to 2.39 hours, indicating that direct spherical sampling improves both geometric correspondence and computational efficiency.

[0068] Table 2 Feature quality of Box RoI and Sphere RoI in different latitudinal ranges low latitude Box RoI 0.43 0.13 0.35 low latitude Sphere RoI 0.48 0.12 0.56 Mid-latitude Box RoI 0.40 0.10 0.31 Mid-latitude Sphere RoI 0.47 0.09 0.54 High latitude Box RoI 0.31 0.12 0.36 High latitude Sphere RoI 0.40 0.07 0.42 overall Box RoI 0.43 0.13 0.34 overall Sphere RoI 0.48 0.12 0.55 Table 2 shows that Sphere RoI outperforms Box RoI at all latitudes, with mIoU increasing from 0.31 to 0.40 in high latitudes and positional deviation decreasing from 0.12 to 0.07. This result corresponds to the latitude-dependent distortion of equidistant cylindrical projection, indicating that local tangent plane sampling and spherical backprojection are not equivalent replacements for ordinary RoIAlign, but rather produce a more significant improvement in geometric distortion in mid-to-high latitude regions.

[0069] Table 3 Ablation results of spherical modeling components and sampling density — √ 1×1 6.87 23.43 √ √ 1×1 7.26 25.16 √ — 2×2 8.33 27.11 √ √ 2×2 9.56 31.04 √ √ 3×3 9.60 30.08 In Table 3, after adding spherical weighting at a 1×1 sampling density, the mAP increased from 6.87% to 7.26%; after adding spherical jitter at a 2×2 sampling density, the mAP increased from 8.33% to 9.56%, and the AP50 increased from 27.11% to 31.04%. Increasing the sampling density from 1×1 to 2×2 significantly improved the positioning quality; although 3×3 slightly increased the mAP to 9.60%, the AP50 decreased to 30.08%, and the number of sampling points increased quadratically. Therefore, this embodiment uses a 2×2 sampling density.

[0070] Table 4 Comparison of the maximum number of iterations T and the convergence threshold ε in spherical weighted spherical scalar ...

[0071] Table 4 shows that the number of iterations and the convergence threshold jointly affect the convergence rate and computation time of spherical weighting. A threshold that is too wide results in limited geometric accuracy, while a threshold that is too tight requires more iterations but the benefits are disproportionate. This embodiment uses... , It achieves a convergence rate of 97.9% with an average of 8.51 actual iterations, balancing the accuracy of spherical center calculation and training efficiency.

[0072] Table 5 Ablation results of scale-rotation fraction consistency loss Not using SRC loss 0.45 9.07 28.85 Using SRC loss 0.48 9.56 31.04 In Table 5, after adding the scale-rotation fractional consistency loss, mIoU increased from 0.45 to 0.48, mAP increased from 9.07% to 9.56%, and AP50 increased from 28.85% to 31.04. Compared to the model without the loss, mAP and AP50 improved by 5.4% and 7.6%, respectively, indicating that the fractional consistency constraint of the corresponding candidate proposals provides transformation-aware supervision that differs from ordinary stochastic augmentation.

[0073] Example 2 A computer device includes a processor and a memory, characterized in that the memory stores program instructions, parameters of a trained pseudo-spherical boundary field of view generation network, and parameters of a panoramic target detector. When the program instructions are executed by the processor, a single-point supervised panoramic target detection method based on spherical geometry modeling, as described in Example 1, is executed.

[0074] Preferably, in the pseudo-spherical boundary field of view label generation stage, the processor loads the pseudo-spherical boundary field of view generation network parameters and generates pseudo-spherical boundary field of view labels based on the equidistant cylindrical projection panoramic image and single-point annotations; in the panoramic target detection deployment stage, the processor loads the panoramic target detector parameters, takes the equidistant cylindrical projection panoramic image to be detected as input, and outputs the target category and the corresponding spherical boundary field of view.

[0075] It is understood that the computer device provided in this embodiment has the same beneficial effects as the single-point supervised panoramic target detection method based on spherical geometry modeling in Embodiment 1, and will not be described again here.

[0076] Example 3 A computer-readable storage medium having a computer program stored thereon, characterized in that, when executed by a processor, the computer program implements a single-point supervised panoramic target detection method based on spherical geometry modeling as described in Example 1.

[0077] It is understood that the computer-readable storage medium provided in this embodiment has the same beneficial effects as the single-point supervised panoramic target detection method based on spherical geometry modeling in Embodiment 1, and will not be repeated here.

[0078] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0079] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A single-point supervised panoramic target detection method based on spherical geometry modeling, characterized in that, include: S1. Initial spherical boundary field of view candidate proposal generation: Obtain the isometric cylindrical projection panoramic image of the original view and the single-point annotation of the target in the isometric cylindrical projection panoramic image, convert the single-point annotation into a spherical position on a unit sphere, and generate an initial spherical boundary field of view candidate proposal with the spherical position as the candidate proposal center. S2. Coarse-grained field of view proposal generation: Based on S1, scale-rotation enhancement is performed on the original view and its initial spherical boundary field of view candidate proposals to obtain a scale-rotation enhanced view and enhanced candidate proposals that correspond one-to-one with the initial spherical boundary field of view candidate proposals. Multi-scale feature extraction is performed on the original view and the scale-rotation enhanced view, respectively. For the initial spherical boundary field of view candidate proposals and the enhanced candidate proposals, spherical region of interest features are extracted based on local tangent plane dense sampling, spherical back projection, and feature pyramid hierarchical allocation. The spherical region of interest features are input into a shared multi-instance learning head for scoring, and scale-rotation score consistency constraints are applied to the scoring results of the initial spherical boundary field of view candidate proposals and the enhanced candidate proposals. Top-k spherical weighting is performed based on the final scores of the initial spherical boundary field of view candidate proposals in the original view to obtain a coarse spherical boundary field of view. S3. Fine field of view proposal optimization: Based on S2, spherical jittering, multi-scale positive candidate proposal generation, negative candidate proposal generation and iterative optimization are performed according to the coarse spherical boundary field of view to obtain pseudo spherical boundary field of view labels. A panoramic target detector is trained using the pseudo-spherical boundary field of view labels; the isometric cylindrical projection panoramic image to be detected is input into the panoramic target detector, and the target category and the corresponding spherical boundary field of view are output. The total loss function for training the pseudo-spherical boundary field of view generation network is as follows: ； in, and The coarse-grained multi-instance learning loss is used for the original view and the scale-rotation augmented view, respectively. For the fine-grained stage of multi-instance learning loss, For negative proposal losses, For scale-rotation fractional consistency loss, and This is the balance coefficient.

2. The single-point supervised panoramic target detection method based on spherical geometry modeling according to claim 1, characterized in that, The generation of initial spherical boundary field of view candidate proposals in S1 specifically includes: The single-point annotation is converted into a spherical position on a unit sphere, and the spherical position corresponding to the single-point annotation is used as the candidate proposal center. A local tangent plane is established at the candidate proposal center, and a minimum proposal scale is set with both the horizontal and vertical field of view angles of 6°. Multi-scale candidate regions are generated on the local tangent plane, and the multi-scale candidate regions are back-projected onto the unit sphere to obtain the spherical boundary field of view candidate proposal set of the original view. The spherical boundary field of view candidate proposal set This is the initial set of candidate proposals for the spherical boundary field of view.

3. The single-point supervised panoramic target detection method based on spherical geometry modeling according to claim 1, characterized in that, The scale-rotation enhancement and multi-scale feature extraction processes in S2 specifically include: The original view is subjected to a planar scale transformation, and the same three-dimensional spherical rotation is applied to the original view and its initial spherical boundary field of view candidate proposals to obtain a scale-rotation enhanced view and a set of enhanced candidate proposals that correspond one-to-one with the initial spherical boundary field of view candidate proposals. The original view and the scale-rotation augmented view are respectively input into the backbone network and the feature pyramid network with shared weights to obtain the multi-scale features of the original view and the multi-scale features of the scale-rotation augmented view.

4. The single-point supervised panoramic target detection method based on spherical geometry modeling according to claim 1, characterized in that, The spherical interest feature extraction process in S2 specifically includes: Based on the spherical geometric parameters of each spherical boundary field of view candidate proposal, a rotatable dense sampling mesh is established on the corresponding local tangent plane; a rotated spherical boundary field of view is then used. Define a local sampling region, where, and Indicates the center of the sphere. and Indicates the horizontal and vertical field of view. Indicates the rotation angle within the tangent plane; when When the rotating spherical boundary field of view degenerates into the spherical boundary field of view, the rotating spherical boundary field of view is only used to define the local sampling region, and the final generated pseudo label is the spherical boundary field of view; At the center of the field of view of the boundary of the rotating spherical surface Construct a central unit vector and tangential basis vectors along local parallels of latitude and longitude. and Let the horizontal half-range of the tangent plane Vertical half range Set the sampling grid resolution. , The sampling density, and ;No. The coordinates of the tangent plane of each sampling point are: ； The first The coordinates of the tangent plane at each sampling point rotate around the center of the candidate proposals in the spherical boundary field of view. : ； Back-project the rotated sampling points onto the unit sphere: ； According to spherical coordinates Obtain the continuous pixel coordinates on the isometric cylindrical projection feature map of the corresponding view; the geometric scale of the spherical boundary field of view candidate proposals. and feature pyramid hierarchy They are respectively: ； ； in, The width of the isometric cylindrical projection panoramic image of the current processing view at the input scale; For reference scale; As the baseline feature level; from the first The layered feature pyramid layer uses bilinear interpolation to extract densely sampled features and aggregates these features into... Features of the region of interest on a sphere.

5. The single-point supervised panoramic target detection method based on spherical geometry modeling according to claim 4, characterized in that, The scale-rotation fraction consistency constraints in S2 specifically include: The initial spherical boundary field of view candidate proposals of the original view and their corresponding augmentation candidate proposals in the scale-rotation augmented view are input into a shared multi-instance learning head to obtain instance scores and class scores; for any view, the first... One candidate proposal, multi-instance learning head output instance score and category scores The final score for the candidate proposals is : ； in, Represents element-wise product; The original view candidate proposal group and its scale-rotation augmented view candidate proposal group corresponding to the same single-point annotation are regarded as a corresponding candidate proposal group. Within each corresponding candidate proposal group, the cosine distance between the initial spherical boundary field of view candidate proposal in the original view and the augmented candidate proposal in the scale-rotation augmented view is calculated as follows: ； in, Indicates a single-point annotation. Indicates the candidate proposal group, This represents the candidate proposals in the original view; Indicates scale - rotation enhancement view and Corresponding enhancement candidate proposals; Represents a fraction type and includes instance fraction types. and category score type The cosine distance is minimized using a smooth L1 loss: ； in, Indicates smoothed L1 loss, and This indicates the balancing weight.

6. The single-point supervised panoramic target detection method based on spherical geometry modeling according to claim 1, characterized in that, The Top-k spherical weighting in S2 specifically includes: Based on the final score of the initial spherical boundary field of view candidate proposals in the original view Select the top-k candidate proposals; normalize the scores of the top-k candidate proposals into weights. ,in Indicates the first The normalized weights corresponding to the Top-k candidate proposals; the spherical centers of each Top-k candidate proposal are converted into three-dimensional unit vectors. ,in Indicates the first The unit vector representation of the spherical center of each of the top-k candidate proposals in three-dimensional Euclidean space. Represents the unit sphere in three-dimensional Euclidean space; Using the weighted sum of normalized three-dimensional unit vectors as the initial center : ； ； Iterative updates are performed using exponential and logarithmic mappings: ； in, Indicated by Spherical exponential mapping with reference to the base point; Indicated by Mapping of the number of spherical surfaces with reference to the base point; Indicates the step size; when the spherical distance between two consecutive iterations is less than the threshold. Or reach the maximum number of iterations Stop at time, and obtain the weighted spherical center. ; The horizontal and vertical field of view of the candidate proposals are weighted according to the following formula: ； in, and They represent the first Horizontal and vertical field of view of the Top-k candidate proposals and These represent the converged horizontal and vertical field of view, respectively; centered on the weighted sphere. The converted latitude and longitude centers and and Together they form a rough spherical boundary field of view.

7. The single-point supervised panoramic target detection method based on spherical geometry modeling according to claim 1, characterized in that, S3 specifically includes: Using the coarse spherical boundary field of view as the initial value, the angular displacement is determined according to the horizontal and vertical field of view angles of the coarse spherical boundary field of view. The spherical surface is shaken in four directions along the lateral great circle direction and the longitudinal meridian direction. Multi-scale positive candidate proposals are generated at each shaking center, and negative candidate proposals with a spherical intersection-union ratio of less than 0.3 are generated. Given the center of the field of view of the spherical boundary Horizontal field of view Vertical field of view and the ratio of shaking ,make: ； When vibrating along the horizontal large circle: ； When vibrating along the longitudinal meridian: ； in, and This indicates the latitude and longitude of the center of the sphere after shaking. When the center of the sphere after shaking crosses the poles, boundary processing is performed. Positive and negative candidate proposals are processed through a shared spherical region of interest feature extraction module and a multi-instance learning head. The prediction scores of positive candidate proposals are normalized and used as spherical weights to update the spherical boundary field of view center and field of view angle. The responses of negative candidate proposals are suppressed by negative proposal loss. The updated spherical boundary field of view generates spherical jitter and multi-scale candidate proposals again. The candidate proposal generation, scoring and spherical weighting process is repeated to obtain the final pseudo-spherical boundary field of view label. The panoramic target detector is trained using the pseudo-spherical boundary field of view label.

8. A computer device comprising a processor and a memory, characterized in that, The memory stores program instructions, parameters of the trained pseudo-spherical boundary field of view generation network, and parameters of the panoramic target detector. When the program instructions are executed by the processor, a single-point supervised panoramic target detection method based on spherical geometric modeling as described in any one of claims 1-7 is executed.

9. A computer device according to claim 8, characterized in that, In the pseudo-spherical boundary field of view label generation stage, the processor loads the pseudo-spherical boundary field of view generation network parameters and generates pseudo-spherical boundary field of view labels based on the equidistant cylindrical projection panoramic image and single-point annotation. During the panoramic target detection deployment phase, the processor loads the panoramic target detector parameters, takes the isometric cylindrical projection panoramic image to be detected as input, and outputs the target category and the corresponding spherical boundary field of view.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a single-point supervised panoramic target detection method based on spherical geometry modeling as described in any one of claims 1-7.

Images