A cross-scale point cloud completion system and method based on probability distribution generation

Abstract

A cross-scale point cloud completion system and method based on probability distribution generation. The system includes a neighborhood distribution feature module, a skeleton point feature module, a cross-scale generation module, and a snowflake deconvolution module. The neighborhood distribution feature module is used to extract features from the obtained original defective point cloud data to obtain the inter-point neighborhood distribution features of the original defective point cloud data. The skeleton point feature module is used to extract features from the sparse defective point cloud data to obtain skeleton point features. The cross-scale generation module is used to perform scale fusion processing and decoding on the inter-point neighborhood distribution features and the skeleton point features to obtain coarsely decoded point cloud data. The snowflake deconvolution module is used to perform upsampling processing on the coarsely decoded point cloud data to obtain the target point cloud data. The present invention has the characteristics of good completion effect for edge loss in the motion state and good completion effect for point cloud targets at different scales.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and particularly relates to a cross-scale point cloud completion system and method based on probability distribution generation. Background Art

[0002] In space on-orbit service and exploration activities, due to the lack of information interaction between space non-cooperative targets and service spacecraft, the risk of vehicle collision is increased. Therefore, the identification and three-dimensional structure perception of space non-cooperative targets are particularly important, and how to effectively identify space non-cooperative targets has become a crucial space mission. Currently, due to the complex environment in space, the uncontrolled motion orbits and poses of non-cooperative targets, and the missing or incomplete point cloud data collected due to the characteristics of the sensors themselves. Under the influence of these factors, the accuracy of satellite identification decreases. Therefore, it is necessary to perform completion processing on the point cloud data collected by the sensors to enhance its three-dimensional structure information. It can be said that the point cloud completion task is an important prerequisite for the completion of non-cooperative target identification and subsequent on-orbit service tasks.

[0003] Traditional point cloud completion methods are usually divided into methods based on geometric structure information and methods based on retrieval templates. The former needs to analyze the surface and spatial structure of the three-dimensional shape to establish manual features for point cloud completion, but such methods are often difficult to extract feature information when dealing with point clouds with a large missing scale, thus limiting their effects in practical applications; while the latter uses known template shapes to estimate the missing or incomplete structure by matching local regions in the point cloud, but for uncataloged space non-cooperative targets, due to the lack of prior model information, it is difficult to achieve its shape completion task based on the retrieval template method.

[0004] Currently, in the deep learning point cloud completion method, by learning prior data, geometric information can be extracted from incomplete inputs to predict complete point clouds, which can recover the detailed features of the three-dimensional shape to a certain extent, thus making up for the shortcomings of traditional point cloud completion methods. However, when the current deep learning-based point cloud completion method faces dynamic targets for point cloud completion, there will still be discontinuous or incomplete point distributions on the surface edge regions of the generated point clouds, resulting in the problem of missing point cloud edges. Summary of the Invention

[0005] In order to overcome the above problems existing in the prior art, the purpose of the present invention is to provide a cross-scale point cloud completion system and method based on probability distribution generation, which has the characteristics of good edge loss completion effect for moving states and good point cloud target completion effect for different scales.

[0006] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0007] A cross-scale point cloud completion system based on probability distribution generation, including a neighborhood distribution feature module, a skeleton point feature module, a cross-scale generation module, and a snowflake deconvolution module;

[0008] The neighborhood distribution feature module is used to extract features from the acquired original defective point cloud data to obtain the inter-point neighborhood distribution features of the original defective point cloud data;

[0009] The skeleton point feature module is used to extract features from the sparse defective point cloud data to obtain skeleton point features, where the sparse defective point cloud data is obtained by downsampling the acquired original defective point cloud data through the farthest point sampling algorithm;

[0010] The cross-scale generation module is used to perform scale fusion processing and decoding on the inter-point neighborhood distribution features and the skeleton point features to obtain the coarsely decoded point cloud data;

[0011] The snowflake deconvolution module is used to perform upsampling on the coarsely decoded point cloud data to obtain the target point cloud data.

[0012] The neighborhood distribution feature module includes a first set abstraction layer, a prior reasoning layer, a fully connected layer, and a first Transformer layer, and the skeleton point feature module includes a second set abstraction layer and a second Transformer layer;

[0013] Through the neighborhood distribution feature module, features are extracted from the original defective point cloud data to obtain the inter-point neighborhood distribution features of the original defective point cloud data;

[0014] Through the first set abstraction layer, features are extracted from the original defective point cloud data to obtain the first high-dimensional features;

[0015] Through the first Transformer layer, the first high-dimensional features are encoded to obtain the first encoded features;

[0016] Through the first set abstraction layer, features are extracted from the first encoded features to obtain the first high-dimensional encoded features;

[0017] Through the prior reasoning layer, the first high-dimensional encoded features are feature-transformed to obtain Gaussian distribution features;

[0018] Through the fully connected layer, the Gaussian distribution features are feature-transformed to obtain the inter-point neighborhood distribution features of the original defective point cloud data.

[0019] Through the skeleton point feature module, features are extracted from the sparse defective point cloud data to obtain skeleton point features;

[0020] Through the second set abstraction layer, features are extracted from the sparse defective point cloud data to obtain the second high-dimensional features;

[0021] Encode the second high-dimensional feature through the second Transformer layer to obtain a second encoded feature;

[0022] Extract features from the second encoded feature through the second set abstraction layer to obtain a second high-dimensional encoded feature;

[0023] Encode the second high-dimensional encoded feature through the second Transformer layer to obtain a double-encoded feature;

[0024] Extract features from the double-encoded feature through the second set abstraction layer to obtain skeleton point features.

[0025] Encode the second high-dimensional feature through the second Transformer layer to obtain a second encoded feature, expressed as the following formula:

[0026]

[0027] where, Δh k-1 represents the second encoded feature, Tran k represents the k-th layer, ΔP in k-1 represents the second high-dimensional feature, represents the spatial position of the skeleton point at the (k - 1)-th layer.

[0028] Extract features from the second encoded feature through the second set abstraction layer to obtain a second high-dimensional encoded feature, expressed as the following formula:

[0029]

[0030] where, ΔP in k represents the second high-dimensional encoded feature, represents the spatial position of the skeleton point at the k-th layer, SA k represents the second set abstraction layer at the k-th layer, Δh k-1 represents the second encoded feature.

[0031] Through the cross-scale generation module, perform scale fusion processing and decoding on the inter-point neighborhood distribution feature and the skeleton point feature to obtain rough decoded point cloud data, including:

[0032] Through the cross-scale generation module, perform scale fusion processing on the inter-point neighborhood distribution feature and the skeleton point feature to generate a global-scale shape feature;

[0033] Decode the global-scale shape feature to obtain rough decoded point cloud data, where the data scale of the rough decoded point cloud data is the same as that of the original defective point cloud data.

[0034] A cross-scale point cloud completion method based on probability distribution generation for the described system, comprising the following steps;

[0035] Step 1: Taking the incomplete point cloud data as input, the distribution feature and the skeleton point feature are respectively obtained by the parallel Neighborhood Distribution Feature Module (NDFM) and Skeleton Point Feature Module (SPSM) as intermediate representations. Among them, the dense point cloud at high resolution passes through the Neighborhood Distribution Feature Module (NDFM) to obtain the intermediate feature f d feature, and the incomplete point cloud is processed to obtain the sparse point cloud at low resolution to pass through the Skeleton Point Feature Module (SPSM) to obtain the intermediate representation f s feature;

[0036] The intermediate representation is then passed through the Cross-scale Generation Module (GCM) to perform scale fusion to generate a rough point cloud structure at different scales, obtaining the global scale shape feature f g , generating a rough point cloud P consistent with the initial scale c ;

[0037] Step 2: The rough point cloud P c is gradually refined into a dense and high-quality point cloud through the multi-level snowflake deconvolution module (SPD), and finally a dense target point cloud that conforms to the real distribution is obtained.

[0038] Specifically, let the input of the incomplete point cloud data P = {p j} be a point cloud set;

[0039] In step 1, when extracting the features of the input point cloud, a parallel structure composed of the Neighborhood Distribution Feature Module (NDFM) and the Skeleton Point Feature Module (SPSM) is adopted. In this structure, the input point cloud set P = {p j} is used as the high-resolution point cloud to be input into the Neighborhood Distribution Feature Module (NDFM) to process the dense point cloud data at high resolution and extract a set of inter-point neighborhood distribution features f d , and the input point cloud set P = {p j} is sampled by the farthest distance sampling algorithm to obtain P l = {p i} as the low-resolution point cloud to be input into the Skeleton Point Feature Module (SPSM) to process the sparse point cloud data at low resolution and capture the skeleton point features f s .

[0040] In the Cross-scale Generation Module (CGM), the inter-point neighborhood distribution feature f d and the skeleton point feature f s are subjected to scale fusion to obtain the global scale shape feature f g, Subsequently, based on the obtained global-scale shape features, a rough point cloud P consistent with the initial scale is generated c , with a size of N c ×3.

[0041] In step 1, in the Skeleton Point Feature Module (SPSM), at the k-th layer, the neighborhood feature representation ΔP of the skeleton points is learned by the set abstraction layer of the k - 1 layer in k-1 and the current spatial position of the skeleton points Take ΔP in k-1 and as inputs and send them to the Transformer module of the K-th layer to enhance the network's understanding ability. ΔP in k-1 is used as the key and query of self-attention, and a new set of local features Δh is generated through the MLP layer and element operations, as shown in the following formula; k-1 , as shown below;

[0042]

[0043] At this time, Δh k-1 is used as the input of the k-th layer set abstraction set, and the input ΔP of the k + 1 layer is obtained in k , as shown below;

[0044]

[0045] In step 2, the SPD module is used to process the point cloud data P obtained in step 1 c , under the guidance of the global-scale shape feature f g , the point cloud data is gradually improved through the multi-level snowflake deconvolution module (SPD) and upsampling factors (represented by r1, r2, r3), and finally the target point cloud Y3 that conforms to the real distribution and is dense is obtained.

[0046] Advantages of the present invention:

[0047] The system and method of the present invention extract features through the neighborhood distribution feature module and the skeleton point feature module, obtaining the inter-point neighborhood distribution features and skeleton point features containing rich point cloud information. Through the cross-scale generation module, the inter-point neighborhood distribution features and skeleton point features are subjected to scale fusion processing to obtain the coarsely decoded point cloud data containing multi-scale point cloud information. Finally, through the snowflake deconvolution module, the coarsely decoded point cloud data is upsampled to obtain the target point cloud data containing complete point cloud information, realizing the completion of point cloud data at multiple scales and the completion of point cloud data under dynamic targets. Since the target point cloud data contains complete point cloud information, the surface edge area of the point cloud generated based on this target point cloud data is relatively smooth, and there is no missing edge in the point cloud. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic flow chart of the present invention.

[0049] Figure 2 It is a schematic diagram of the dual-channel encoding method.

[0050] Figure 3 It is a schematic diagram of the cross-scale generation module.

[0051] Figure 4 It is a schematic diagram of the cross-scale point cloud completion process of the cross-scale point cloud completion method based on probability distribution in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] The present invention will be further described in detail below with reference to the accompanying drawings.

[0053] See Figure 1 , the overall network structure of the cross-scale point cloud completion system based on probability distribution consists of a neighborhood distribution feature module (NDFM), a skeleton point feature module (SPSM), a cross-scale generation module (CGM), and a snowflake deconvolution module (SPD). In the first stage, the incomplete point cloud data is used as the input. The parallel NDFM and SPSM respectively obtain their distribution features and skeleton point features as intermediate representations, and then the cross-scale generation module (GCM) is used to perform scale fusion to generate rough point cloud structures at different scales. In the second stage, through the multi-level snowflake deconvolution module (SPD), a dense and high-quality point cloud is further restored from the complete but rough point cloud.

[0054] Specifically, let the input P = {p j} be a point cloud set. In the first stage, when extracting the features of the input point cloud, a parallel structure composed of a neighborhood distribution feature module (NDFM) and a skeleton point feature module (SPSM) is adopted. In this structure, the neighborhood distribution feature module (NDFM) focuses on processing dense point cloud data and extracting a set of inter-point neighborhood distribution features f d, while the Skeleton Point Feature Module (SPSM) specifically captures the skeleton point feature f of the overall point cloud object for sparse point cloud data s .

[0055] In the Cross-scale Generation Module (CGM), f d and f s are subjected to scale fusion to obtain the global-scale shape feature f g containing the probability distribution information of the original point cloud. Subsequently, based on the obtained global-scale shape feature, a rough point cloud P c consistent with the initial scale is generated, with a size of N c ×3. In the second stage, the Snowflake Deconvolution Module (SPD) is used to process the point cloud data P c obtained in the first stage. Guided by the global-scale shape feature f g , the point cloud data is gradually refined through the multi-level Snowflake Deconvolution Module (SPD) and the upsampling factors (represented by r1, r2, r3), and finally the target point cloud Y3 that conforms to the real distribution and is dense is obtained.

[0056] As Figure 2 shown, in the dual-channel encoding, the Neighborhood Distribution Feature Module (NDFM) and the Skeleton Point Feature Module (SPSM) respectively use high-resolution and low-resolution point clouds as inputs to obtain the inter-point neighborhood distribution feature f d and the skeleton point feature f s .

[0057] In the Neighborhood Distribution Feature Module (NDFM), the original point cloud P = {p j} is used as the high-resolution channel input, and the neighborhood features of the skeleton points are obtained through the Set Abstraction layer (SA); subsequently, the relationships between points are globally modeled and encoded through the Transformer module, and the neighborhood features obtained by the SA layer are further learned.

[0058] Through the Prior Inference (PI) layer, the neighborhood feature representation is losslessly transformed into the standard Gaussian distribution, and then the data that conforms to the real probability distribution is screened through the random sampling operation;

[0059] Finally, the inter-point neighborhood distribution feature f d containing the original probability distribution information is generated through the FC layer (FC).

[0060] In the Skeleton Point Feature Module (SPSM), the input point cloud set P = {p j} is downsampled using the farthest point sampling algorithm to obtain a sparse point cloud set P l = {p i} as the input; the main purpose of the SPSM module is to extract the bone point features f in the overall three-dimensional shape s ;

[0061] In PointNet++, the set abstraction layer (SA) is used to extract the bone points of the sparse point cloud and their spatial position information as the bone point features of the overall shape. At the same time, the Transformer module is added to enable the SPSM module to more effectively learn the relationship between the bone point features and their spatial positions contained in the overall shape.

[0062] Specifically, at the k-th layer, the neighborhood feature representation ΔP of the bone points is learned by the set abstraction layer of the k-1 layer in k-1 and the current bone point spatial position Take ΔP in k-1 and as the input to the Transformer module of the K-th layer to enhance the network's understanding ability. ΔP in k-1 as the key and query of self-attention, and a new set of local features Δh is generated through a series of MLP layers and element operations k-1 . As shown in the following formula.

[0063]

[0064] At this time, Δh k-1 as the input of the k-th layer set abstraction set, and the input ΔP of the k+1 layer is obtained in k , As shown in the following formula.

[0065]

[0066] As Figure 3 shown: The cross-scale generation module (CGM) consists of a scale fusion layer, a transposed convolution layer, and an MLP. The overall structure is as Figure 3 shown. The features f d and f s obtained using the parallel structure are combined through scale fusion to obtain the global scale shape feature f g . Subsequently, a segmentation operation is performed through transposed convolution to generate point features point by point to capture the existing and missing structures. Each point feature is aggregated with the shape feature through a multi-layer perceptron to generate the point cloud P co . Finally, P co is merged with the input point cloud P = {p j}, and the merged point cloud is downsampled to P c using the farthest point sampling algorithm.

[0067] The Chamfer distance (CD) is selected as the loss function to balance computational efficiency and accuracy. The Chamfer distance (CD) under the l2 norm is defined as follows:

[0068]

[0069] where in the formula, P and Q represent two sets of point clouds respectively. The first term represents the sum of the minimum distances from any point x in P to Q, and the second term represents the sum of the minimum distances from any point y in Q to P.

[0070] At the same time, multi-level losses are adopted for guidance, as shown below:

[0071]

[0072] where P c ′ and Y i ′ are respectively obtained by downsampling the corresponding P c , Y i through the real point cloud and have the same number of point clouds.

[0073] The loss function is like a ruler in the whole algorithm process, used to measure the gap between the model prediction result and the real result. In machine learning and deep learning, the goal of this application is to train a model to accurately predict unknown data, and the loss function helps to evaluate the accuracy of the model prediction.

[0074] When training the model with training data, the loss function calculates a value based on the model's prediction result and the real label. This value represents the quality of the model's prediction under the current parameters. Then, through an optimization algorithm (such as gradient descent), the model's parameters are continuously adjusted to minimize the value of the loss function, that is, to make the model's prediction result as close as possible to the real result.

[0075] Therefore, the loss function can guide the update direction of the model parameters, help the model to be continuously optimized, improve the prediction accuracy, and thus enable the machine learning model to better complete the task.

[0076] Example: As Figure 4 the input defective point cloud set P = {p j}, the defective point cloud set is set as the high-resolution point cloud P = {p j}, input to the neighborhood distribution feature module (NDFM), and the defective point cloud set is sampled by the farthest point sampling algorithm to obtain the low-resolution point cloud P l = {p i},

[0077] input to the skeleton point feature module (SPSM), and the obtained features fd With f s perform scale fusion to obtain f g Based on f g generate a rough point cloud P consistent with the initial scale c , and finally process the point cloud data P through the snowflake deconvolution module (SPD) c , under the guidance of the global-scale shape feature f g , gradually improve the point cloud data through the multi-level snowflake deconvolution module (SPD) and upsampling factors (represented by r1, r2, r3), and finally obtain the target point cloud Y3 that conforms to the real distribution and is dense.

[0078] The present invention also includes an electronic device, including a memory and a processor. The memory is used to store various computer program instructions, and the processor is used to execute the computer program instructions to complete all or part of the above steps; the electronic device can communicate with one or more external devices, and can also communicate with one or more devices that enable a user to interact with the electronic device, and / or communicate with any device that enables the electronic device to communicate with one or more other computing devices. The electronic device can also communicate with one or more networks (such as local area networks, wide area networks, and / or public networks) through a network adapter.

[0079] The present invention also includes a computer-readable storage medium storing a computer program, which can be executed by a processor. The computer-readable storage medium can include, but is not limited to, magnetic storage devices, optical discs, digital versatile discs, smart cards, and flash memory devices. In addition, the readable storage medium described in the present invention can represent one or more devices and / or other machine-readable media for storing information. The term "machine-readable medium" includes, but is not limited to, wireless channels and various other media (and / or storage media) that can store, contain, and / or carry code and / or instructions and / or data.

Claims

1. A cross-scale point cloud completion system generated based on probability distribution, characterized in that, It includes a neighborhood distribution feature module, a skeleton point feature module, a cross-scale generation module, and a snowflake deconvolution module; The neighborhood distribution feature module is used to extract features from the obtained original defective point cloud data to obtain the inter-point neighborhood distribution features of the original defective point cloud data; The skeleton point feature module is used to extract features from the sparse defective point cloud data to obtain skeleton point features, where the sparse defective point cloud data is obtained by downsampling the obtained original defective point cloud data through the farthest point sampling algorithm; The cross-scale generation module is used to perform scale fusion processing and decoding on the inter-point neighborhood distribution features and skeleton point features to obtain coarsely decoded point cloud data; The snowflake deconvolution module is used to upsample the coarsely decoded point cloud data to obtain the target point cloud data; The neighborhood distribution feature module includes a first set abstraction layer, a prior reasoning layer, a fully connected layer, and a first Transformer layer; The skeleton point feature module includes a second set abstraction layer and a second Transformer layer; The neighborhood distribution feature module extracts features from the original defective point cloud data to obtain the inter-point neighborhood distribution features of the original defective point cloud data; The first set abstraction layer extracts features from the original defective point cloud data to obtain a first high-dimensional feature; The first Transformer layer encodes the first high-dimensional feature to obtain a first encoded feature; The first set abstraction layer extracts features from the first encoded feature to obtain a first high-dimensional encoded feature; The prior reasoning layer performs feature transformation on the first high-dimensional encoded feature to obtain a Gaussian distribution feature; The fully connected layer performs feature transformation on the Gaussian distribution feature to obtain the inter-point neighborhood distribution features of the original defective point cloud data; The skeleton point feature module extracts features from the sparse defective point cloud data to obtain skeleton point features; The second set abstraction layer extracts features from the sparse defective point cloud data to obtain a second high-dimensional feature; The second Transformer layer encodes the second high-dimensional feature to obtain a second encoded feature; The second set abstraction layer extracts features from the second encoded feature to obtain a second high-dimensional encoded feature; The second Transformer layer encodes the second high-dimensional encoded feature to obtain a double-encoded feature; The second set abstraction layer extracts features from the double-encoded feature to obtain skeleton point features.

2. The cross-scale point cloud completion system generated based on probability distribution according to claim 1, wherein, The second Transformer layer encodes the second high-dimensional feature to obtain a second encoded feature, expressed as the following formula: Among them, Δh k-1 represents the second coding feature, Tran k represents the k-th layer, ΔP in k-1 represents the second high-dimensional feature, represents the spatial position of the bone point at the (k - 1)-th layer.

3. The cross-scale point cloud completion system generated based on probability distribution according to claim 1, characterized in that, The second set abstraction layer extracts features from the second encoded feature to obtain a second high-dimensional encoded feature, expressed as the following formula: Among them, ΔP in k represents the second high-dimensional encoded feature, represents the spatial position of the bone point at the k-th layer, SA k represents the second set abstraction layer at the k-th layer, Δh k-1 represents the second encoded feature.

4. A cross-scale point cloud completion system generated based on probability distribution according to claim 1, characterized in that, The cross-scale generation module performs scale fusion processing and decoding on the inter-point neighborhood distribution features and skeleton point features to obtain coarsely decoded point cloud data; The cross-scale generation module performs scale fusion processing on the inter-point neighborhood distribution features and skeleton point features to generate global-scale shape features; The global-scale shape features are decoded to obtain coarsely decoded point cloud data, where the data scale of the coarsely decoded point cloud data is the same as that of the original defective point cloud data.

5. A cross-scale point cloud completion method based on probability distribution for the system according to any one of claims 1-4, characterized in that, It includes the following steps: Step 1: Take the incomplete point cloud data as input. The parallel neighborhood distribution feature module and the bone point feature module respectively obtain its distribution feature and bone point feature as intermediate representations. Among them, the dense point cloud at high resolution passes through the neighborhood distribution feature module to obtain the intermediate feature f d feature. The incomplete point cloud is processed to obtain a sparse point cloud at low resolution to pass through the bone point feature module to obtain the intermediate representation f s feature; The intermediate representation is then subjected to scale fusion through a cross-scale generation module to generate a rough point cloud structure at different scales, obtaining the global-scale shape feature f g , generating a rough point cloud P consistent with the initial scale c ; Step 2: Coarse point cloud P c The dense and high-quality point cloud is gradually refined through the multi-level snowflake deconvolution module, and finally the target point cloud that conforms to the real distribution and is dense is obtained.

6. A cross-scale point cloud completion method generated based on probability distribution according to claim 5, characterized in that Let the input of the incomplete point cloud data be \(P = \{p j \}\) be a point cloud set; In step 1, when extracting the input point cloud features, a parallel structure composed of a neighborhood distribution feature module and a skeleton point feature module is adopted. In this structure, the input point cloud set P = {p j} is used as the high-resolution point cloud and input into the neighborhood distribution feature module to process the dense point cloud data at high resolution and extract a set of inter-point neighborhood distribution features f d . The input point cloud set P = {p j} is sampled by the farthest distance sampling algorithm to obtain P l = {p i}, which is used as the low-resolution point cloud and input into the skeleton point feature module to process the sparse point cloud data at low resolution and capture the skeleton point features f s of the overall point cloud object.

7. A cross-scale point cloud completion method generated based on probability distribution according to claim 6, characterized in that In the cross-scale generation module, the inter-point neighborhood distribution feature f d and the bone point feature f s are subjected to scale fusion to obtain the global-scale shape feature f g containing the probability distribution information of the original point cloud. Subsequently, based on the obtained global-scale shape feature, a rough point cloud P c consistent with the initial scale is generated, with a size of N c ×3.

8. A cross-scale point cloud completion method generated based on probability distribution according to claim 6, characterized in that, In the step 1, in the bone point feature module, at the k-th layer, the neighborhood feature representation ΔP of the bone point is learned by the set abstraction layer of the (k - 1)-th layer in k-1 and the current spatial position of the bone point Take ΔP in k-1 and as inputs and send them to the Transformer module of the k-th layer to enhance the network's understanding ability. ΔP in k-1 As the key and query of self-attention, a new set of local features Δh is generated through the MLP layer and element-wise operations, as shown in the following formula; k-1 , as shown in the following formula; At this time, Δh k-1 is used as the input of the k-th layer set abstraction set to obtain the input ΔP of the (k + 1)-th layer in k , as shown in the following formula; 9. A cross-scale point cloud completion method generated based on probability distribution according to claim 5, characterized in that In the said step 2, a snowflake deconvolution module is adopted to process the point cloud data P obtained in step 1 c , under the guidance of the global-scale shape feature f g , the point cloud data is gradually improved through a multi-level snowflake deconvolution module and upsampling factors, and finally the target point cloud Y3 that conforms to the true distribution and is dense is obtained.

Images