A model quantization method, device, system and storage medium for point cloud data

Abstract

The application provides a point cloud data model quantization method, device and system and a storage medium, and relates to the technical fields of financial technology and automatic driving. The point cloud data model quantization method comprises the following steps: inputting point cloud data into a convolutional neural network to extract a channel feature distribution map; performing normalization processing on each channel feature in the channel feature distribution map to obtain a channel feature parameter, and converting the channel feature parameter into an integer point number representation; and performing quantization training on a model based on the channel feature parameter by using a back propagation method to obtain quantization data. The model quantization method provided by the application can effectively reduce the calculation amount of a deep neural network, improve the calculation speed, effectively save the calculation cost and resource cost, and improve the real-time performance of network operation, so that the application is better applied to scenarios such as automatic driving and virtual reality.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and autonomous driving technology, and more specifically, to a method, apparatus, system and storage medium for model quantization of point cloud data. Background Technology

[0002] In recent years, the emergence of concepts such as autonomous driving, high-precision maps, virtual reality, and augmented reality has greatly promoted the development of large-scale 3D point cloud semantic understanding and analysis. In the fintech sector, numerous financial institutions are extensively applying autonomous driving and virtual reality technologies. For example, in the actual business processes of financial institutions, funds, materials, and equipment are automatically transferred between different institutions using transportation, replacing manual escort, transport, and delivery work.

[0003] In technologies such as autonomous driving and virtual reality, 3D point cloud semantic segmentation is the process of classifying each point in a chaotic point cloud by assigning semantic labels, which is a key step in scene understanding.

[0004] Currently, 3D point cloud segmentation is mainly achieved based on deep neural networks. This involves constructing deep neural networks to extract features from point clouds, thereby enabling 3D point cloud segmentation. However, point cloud data typically contains XYZ positional information, RGB color information, and I intensity information, making it a complex, multi-dimensional dataset. Furthermore, the sheer volume of point clouds in real-world scenes necessitates enormous computational overhead, leading to reduced computational efficiency, difficulty in achieving required accuracy, and significant consumption of system resources. This hinders the advancement of technologies such as autonomous driving and virtual reality. Summary of the Invention

[0005] In view of this, and to address the aforementioned technical problems, the present invention provides a method for model quantization of point cloud data, comprising: Point cloud data is input into a convolutional neural network to extract channel feature distribution maps; Normalize each channel feature in the channel feature distribution map to obtain channel feature parameters, and convert the channel feature parameters from floating-point representation to integer representation; Based on the channel feature parameters, the model is trained using backpropagation to obtain quantized data.

[0006] Preferably, the step of inputting point cloud data into a convolutional neural network to extract channel feature distribution maps includes: The point cloud data is divided into a set of data blocks of equal size according to their spatial location; The data block set is input one by one into a convolutional neural network to extract the channel feature distribution map.

[0007] Preferably, the normalization process for each channel feature in the channel feature distribution map to obtain channel feature parameters includes: Each channel feature is subjected to adaptive normalization based on the quantization bit width to obtain the channel feature parameters; The adaptive normalization processing method is as follows: ; Where s(n) is the interval of quantization operations, and n is the quantization bit width.

[0008] Preferably, the adaptive normalization processing based on the quantization bit width for each channel feature further includes: After activating the function layer, the information type of the channel features is determined; If the information type of the channel feature is XYZ position information, then the corresponding quantization bit width is 4 bits; If the information type of the channel feature is RGB color information or I intensity information, then the corresponding quantization bit width is 8 bits.

[0009] Preferably, the step of converting the channel feature parameters from floating-point representation to integer representation includes: The normalized channel feature parameters are discretized, and the corresponding floating-point numbers are mapped to an n-bit integer range, thus converting the channel feature parameters from floating-point representation to integer representation.

[0010] Preferably, the step of quantizing the model using backpropagation based on the channel feature parameters to obtain quantized data includes: The mean and variance obtained from the forward propagation are statistically analyzed, and the output of the convolution is summed along the feature channel dimension of the channel feature parameters. During the summation process, the already saved mean and variance are used to perform precision conversion again to obtain the quantized data.

[0011] Preferably, in the summation process, the precision conversion is performed again using the already stored mean and variance to obtain the quantized data, including: The mean and variance are converted to higher precision to obtain the weights through the quantization training. The weights are saved as 8-bit data as the quantization data.

[0012] Furthermore, to address the aforementioned problems, the present invention also provides a model quantization device for point cloud data, comprising: The extraction module is used to input point cloud data into a convolutional neural network and extract channel feature distribution maps; The normalization module is used to normalize each channel feature in the channel feature distribution map to obtain channel feature parameters, and convert the channel feature parameters from floating-point representation to integer representation; The training module is used to perform quantization training on the model based on the channel feature parameters using backpropagation to obtain quantized data.

[0013] In addition, to solve the above problems, the present invention also provides a model quantization system for point cloud data, including a memory and a processor. The memory stores a model quantization program for point cloud data, and the processor runs the model quantization program for point cloud data to enable the model quantization system for point cloud data to perform the model quantization method for point cloud data as described above.

[0014] In addition, to solve the above problems, the present invention also provides a computer-readable storage medium storing a model quantization program for point cloud data, wherein the model quantization program for point cloud data is executed by a processor to implement the model quantization method for point cloud data as described above.

[0015] This invention provides a method, apparatus, system, and storage medium for model quantization of point cloud data. The method includes: inputting point cloud data into a convolutional neural network to extract a channel feature distribution map; normalizing each channel feature in the channel feature distribution map to obtain channel feature parameters, and converting the channel feature parameters from floating-point representation to integer representation; and performing quantization training on the model using backpropagation based on the channel feature parameters to obtain quantized data. This invention addresses the multi-channel dimensionality of point cloud data by using a quantization method to train the model, thereby obtaining quantized data. Introducing a model quantization method can effectively reduce the computational load of deep neural networks, improve computational speed, effectively save computational and resource costs, and improve the real-time performance of network operation, thus making it better applicable to scenarios such as autonomous driving and virtual reality. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the hardware operating environment involved in an embodiment of the point cloud data model quantization method of the present invention; Figure 2 This is a flowchart illustrating the first embodiment of the point cloud data model quantization method of the present invention; Figure 3 This is a detailed flowchart of step S100 in the second embodiment of the point cloud data model quantization method of the present invention. Figure 4 This is a detailed flowchart of step S200 in the second embodiment of the point cloud data model quantization method of the present invention; Figure 5This is a flowchart illustrating the step of determining the information type in step S200 of the second embodiment of the point cloud data model quantization method of the present invention. Figure 6 This is a schematic diagram of the overall process of step S200 refinement (including the refinement of the integer point number conversion step) in the second embodiment of the point cloud data model quantization method of the present invention; Figure 7 This is a flowchart illustrating the third embodiment of the point cloud data model quantization method of the present invention; Figure 8 This is a detailed flowchart of step 310 in the third embodiment of the point cloud data model quantization method of the present invention. Figure 9 This is a schematic diagram of the module connections of the point cloud data model quantization device of the present invention.

[0017] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] The embodiments of the present invention are described in detail below, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0019] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0022] like Figure 1 The diagram shown is a structural schematic of the hardware operating environment of the terminal involved in an embodiment of the present invention.

[0023] The point cloud data model quantization system of this invention can be a PC, or a mobile terminal device such as a smartphone, tablet, or laptop. This point cloud data model quantization system may include: a processor 1001 (e.g., a CPU), a network interface 1004, a user interface 1003, a memory 1005, and a communication bus 1002. The communication bus 1002 is used to establish communication between these components. The user interface 1003 may include a display screen, an input unit such as a keyboard, or a remote control; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed RAM memory or a stable memory, such as a disk storage device. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001. Optionally, the point cloud data model quantization system may also include RF (Radio Frequency) circuitry, audio circuitry, a Wi-Fi module, etc. In addition, the point cloud data model quantization system can also be configured with other sensors such as gyroscopes, barometers, hygrometers, thermometers, and infrared sensors, which will not be elaborated here.

[0024] Those skilled in the art will understand that Figure 1 The model quantization system for point cloud data shown is not intended to limit it and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. Figure 1 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a data interface control program, a network connection program, and a model quantization program for point cloud data.

[0025] In summary, this invention addresses the multi-channel dimensionality of point cloud data by using a quantization method to train the model, thereby obtaining quantized data. Introducing a model quantization method can effectively reduce the computational load of deep neural networks, improve computational speed, effectively save computational and resource costs, and improve the real-time performance of network operation, thus making it better applicable to scenarios such as autonomous driving and virtual reality.

[0026] Example 1: Reference Figure 2 The first embodiment of the present invention provides a model quantization method for point cloud data, comprising: Step S100: Input the point cloud data into the convolutional neural network to extract the channel feature distribution map; As mentioned above, point cloud data refers to a collection of vectors in a three-dimensional coordinate system. Scanned data is recorded in the form of points, each containing three-dimensional coordinates, and some may contain color information (RGB) or reflectance intensity information.

[0027] As mentioned above, Convolutional Neural Networks (CNNs) are a type of feedforward neural network that includes convolutional computations and has a deep structure. They are one of the representative algorithms of deep learning. CNNs have representation learning capabilities and can perform shift-invariant classification of input information according to their hierarchical structure.

[0028] By using a convolutional neural network, point cloud data is divided and deep learned separately to extract the channel feature distribution map.

[0029] Step S200: Normalize each channel feature in the channel feature distribution map to obtain channel feature parameters, and convert the channel feature parameters from floating-point representation to integer representation; As mentioned above, normalization is a dimensionless processing method that transforms the absolute values ​​of physical system values ​​into relative values. It is an effective way to simplify calculations and reduce the magnitude of quantities.

[0030] A convolutional neural network (CNN) consists of one or more convolutional layers and a fully connected layer at the top (corresponding to a classic neural network), as well as associated weights and pooling layers. Compared to other deep learning architectures, CNNs can deliver better results in image and speech recognition. This model can also be trained using the backpropagation algorithm. Compared to other deep, feedforward neural networks, CNNs can achieve higher performance with fewer parameters.

[0031] Based on convolutional neural networks, feature maps are extracted from the input data set of point cloud data, which are called feature distribution maps.

[0032] The extracted feature distribution map, after normalization, yields channel feature parameters. After data processing, such as discretization and mapping, these channel feature parameters are represented as floating-point numbers. This representation is then converted to integer representation to facilitate further quantization training.

[0033] As mentioned above, floating-point representation is the standard IEEE floating-point representation used by known C / C++ compilers. This structure is a scientific notation that uses a sign (+ or -), an exponent, and a mantissa, with the base being 2. Therefore, in IEEE floating-point representation, a floating-point number is the mantissa multiplied by a power of 2 plus the sign.

[0034] Step S300: Based on the channel feature parameters, the model is quantized and trained using backpropagation to obtain quantized data.

[0035] As mentioned above, backpropagation (BP) is short for "error backpropagation," a common method used in conjunction with optimization methods (such as gradient descent) to train artificial neural networks. This method calculates the gradient of the loss function over all weights in the network. This gradient is fed back to the optimization method to update the weights to minimize the loss function. The main algorithm for performing gradient descent on a neural network is as follows: This algorithm first calculates (and caches) the output value of each node using forward propagation, and then calculates the partial derivative of the loss function with respect to each parameter by traversing the graph using backpropagation.

[0036] By using backpropagation, the model is trained with quantization based on existing or obtained channel feature parameters, thereby obtaining the corresponding quantized data.

[0037] This embodiment takes advantage of the multi-channel and multi-dimensional characteristics of point cloud data and uses a quantization method to quantize and train the model, thereby obtaining quantized data. Introducing a model quantization method can effectively reduce the computational load of deep neural networks, improve computational speed, effectively save computational and resource costs, and improve the real-time performance of network operation, thus making it better applicable to scenarios such as autonomous driving and virtual reality.

[0038] Example 2: Reference Figure 3 The second embodiment of the present invention provides a model quantization method for point cloud data, based on the above embodiment 1. Step S100 involves inputting the point cloud data into a convolutional neural network to extract a channel feature distribution map, including: Step S110: Divide the point cloud data into a set of data blocks of equal size according to their spatial location; As mentioned above, point cloud data contains XYZ position information, RGB color information, and I intensity information, making it a complex multi-dimensional data.

[0039] By using spatial location, the currently acquired point cloud data can be divided according to space, and further data acquisition can be performed.

[0040] Step S120: Input the data block set one by one into the convolutional neural network to extract the channel feature distribution map.

[0041] The 3D point cloud data is divided according to its spatial position, that is, into a set of data blocks of uniform size. The set of point cloud data blocks is then input into a convolutional neural network one by one to extract the channel feature distribution map.

[0042] Each channel feature in the channel feature distribution map is normalized to obtain channel feature parameters, and the channel feature parameters are converted from floating-point representation to integer representation. Further reference Figure 4 Step S200 involves normalizing each channel feature in the channel feature distribution map to obtain channel feature parameters, including: Step S210: Perform adaptive normalization processing based on quantization bit width on each channel feature to obtain the channel feature parameters; The adaptive normalization processing method is as follows: ; Where S(n) is the interval of quantization operations, and n is the quantization bit width.

[0043] As mentioned above, normalization utilizes the maximum and minimum values ​​of features (channel features) to scale the feature values ​​to the [0,1] interval. The min-max function is used for scaling each column of features. Its purpose is to eliminate dimensionality and accelerate convergence: different features often have different units of measurement, which can affect the results of data analysis. To eliminate the influence of dimensions between indicators, data normalization is necessary to ensure comparability between data indicators. After data normalization, the original data contains decimals between [0,1], making it suitable for comprehensive comparative evaluation.

[0044] The above, This is the data after normalization.

[0045] As described above, in the point cloud feature distribution map, different channels all have non-zero means, and the channel means are also unequal among different block sets. Therefore, before performing further quantization operations, it is necessary to use the above calculation method to perform adaptive normalization processing on the channel features of each channel based on the quantization bit width.

[0046] Further reference Figure 5 Step S200, which performs adaptive normalization processing based on the quantization bit width for each channel feature, further includes: Step S220: After the activation function layer, determine the information type of the channel features; The information type mentioned above refers to the data type of point cloud data corresponding to the channel features, including XYZ position information, RGB color information, and I intensity information.

[0047] Step S230: If the information type of the channel feature is XYZ position information, then the corresponding quantization bit width is 4 bits.

[0048] Step S240: If the information type of the channel feature is RGB color information or I intensity information, then the corresponding quantization bit width is 8 bits.

[0049] As described above, different quantization bit widths are selected for different feature channels. Specifically, after the activation function layer, a 4-bit quantization bit width is selected for the features obtained from the XYZ position information, and an 8-bit quantization bit width is selected for the RGB color information and the I intensity information.

[0050] Further reference Figure 6 Step S200, converting the channel feature parameters from floating-point representation to integer representation, includes: Step S250: Discretize the normalized channel feature parameters, then map the corresponding floating-point numbers to an n-bit integer range, and convert the channel feature parameters from floating-point representation to integer representation.

[0051] Discretization, as described above, maps a finite number of individuals in an infinite space to a finite space, thereby improving the time and space efficiency of the algorithm. In other words, discretization reduces the size of the data without altering its relative magnitude.

[0052] Example 3: Reference Figure 7 The third embodiment of the present invention provides a model quantization method for point cloud data. Based on the above embodiment 1, step S300 involves quantizing the model using backpropagation based on the channel feature parameters to obtain quantized data, including: Step S310: Statistically calculate the mean and variance obtained from the forward propagation, and sum the output of the convolution along the feature channel dimension of the channel feature parameters; during the summation process, use the already saved mean and variance to perform precision conversion again to obtain the quantized data.

[0053] As mentioned above, 5. Before performing backpropagation of the neural network, the mean and variance obtained from forward propagation are statistically analyzed, and the output of the convolution is summed along the feature channel dimension. During the summation process, the saved mean and variance are used again for precision conversion to ensure that the network is differentiable.

[0054] Further reference Figure 8In step S310, during the summation process, the already saved mean and variance are used to perform precision conversion again to obtain the quantized data, including: Step S311: The mean and variance are converted to higher precision again to obtain the weights through the quantization training; Step S312: Save the weight as 8-bit wide data as the quantization data.

[0055] The weights obtained after quantization and training are saved as 8-bit data, which can be used for deployment on edge devices such as in-vehicle microcomputers to reduce computing power and thus improve the real-time performance of point cloud segmentation.

[0056] In summary, a 3D point cloud semantic segmentation scenario involves tens of millions to hundreds of millions of unordered 3D point clouds. Each point cloud data contains multiple dimensions of information, including XYZ position information, RGB color information, and I intensity information. Processing these irregular and unordered 3D point clouds using deep neural networks via convolution requires significant computational power and time. Therefore, this embodiment provides a multi-channel perception quantization scheme to address the multi-dimensional characteristics of point cloud data. Specifically, it introduces model quantization to reduce computational costs and improve network speed.

[0057] In addition, refer to Figure 9 This embodiment also provides a model quantization device for point cloud data, including: Extraction module 10 is used to input point cloud data into a convolutional neural network and extract channel feature distribution maps; The normalization module 20 is used to normalize each channel feature in the channel feature distribution map to obtain channel feature parameters, and to convert the channel feature parameters from floating-point representation to integer representation. Training module 30 is used to perform quantization training on the model based on the channel feature parameters using backpropagation to obtain quantized data.

[0058] Furthermore, this embodiment also provides a model quantization system for point cloud data, including a memory and a processor. The memory stores a model quantization program for point cloud data, and the processor runs the model quantization program for point cloud data to enable the model quantization system for point cloud data to perform the model quantization method for point cloud data as described above.

[0059] In addition, this embodiment also provides a computer-readable storage medium storing a model quantization program for point cloud data. When the model quantization program for point cloud data is executed by a processor, it implements the model quantization method for point cloud data as described above.

[0060] In summary, the point cloud data model quantization method provided by this invention utilizes a quantization method to train the model based on the multi-channel dimensional characteristics of point cloud data, thereby obtaining quantized data. Introducing a model quantization method can effectively reduce the computational load of deep neural networks, improve computational speed, effectively save computational and resource costs, and improve the real-time performance of network operation, thus making it better applicable to scenarios such as autonomous driving and virtual reality.

[0061] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0062] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention. The above are only preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for model quantization of point cloud data, characterized in that, include: Point cloud data is input into a convolutional neural network to extract channel feature distribution maps; Each channel feature in the channel feature distribution map is normalized to obtain channel feature parameters, and the channel feature parameters are converted from floating-point representation to integer representation; wherein, the normalization process for each channel feature in the channel feature distribution map to obtain channel feature parameters includes: performing adaptive normalization processing based on the quantization bit width for each channel feature to obtain the channel feature parameters; the method of adaptive normalization processing is as follows: Where s(n) is the interval of quantization operations, and n is the quantization bit width; in the point cloud feature distribution map, different channels have non-zero mean values, and the channel mean values ​​are not equal between different block sets; therefore, before performing further quantization operations, it is necessary to use the above calculation method to perform adaptive normalization processing based on the quantization bit width on the channel features of each channel; the adaptive normalization processing based on the quantization bit width on each channel feature also includes: after the activation function layer, determining the information type of the channel feature; if the information type of the channel feature is XYZ position information, the corresponding quantization bit width is 4 bits; if the information type of the channel feature is RGB color information or I intensity information, the corresponding quantization bit width is 8 bits; Based on the channel feature parameters, the model is quantized and trained using backpropagation to obtain quantized data. This includes: before performing backpropagation of the neural network, the mean and variance obtained from forward propagation are statistically analyzed, the output of the convolution is summed along the feature channel dimension, and the already saved mean and variance are used to perform precision conversion again during the summation process to ensure that the network is differentiable.

2. The point cloud data model quantization method as described in claim 1, characterized in that, The step of inputting point cloud data into a convolutional neural network to extract channel feature distribution maps includes: The point cloud data is divided into a set of data blocks of equal size according to their spatial location; The data block set is input one by one into a convolutional neural network to extract the channel feature distribution map.

3. The point cloud data model quantization method as described in claim 1, characterized in that, The step of converting the channel feature parameters from floating-point representation to integer representation includes: The normalized channel feature parameters are discretized, and the corresponding floating-point numbers are mapped to an n-bit integer range, thus converting the channel feature parameters from floating-point representation to integer representation.

4. The point cloud data model quantization method as described in claim 1, characterized in that, In the summation process, the already saved mean and variance are used to perform precision conversion again to obtain the quantified data, including: The mean and variance are converted to higher precision to obtain the weights through the quantization training. The weights are saved as 8-bit data as the quantization data.

5. A model quantization device for point cloud data, characterized in that, include: The extraction module is used to input point cloud data into a convolutional neural network and extract channel feature distribution maps; The normalization module is used to normalize each channel feature in the channel feature distribution map to obtain channel feature parameters, and to convert the channel feature parameters from floating-point representation to integer representation; wherein, the normalization process for each channel feature in the channel feature distribution map to obtain channel feature parameters includes: performing adaptive normalization processing based on the quantization bit width for each channel feature to obtain the channel feature parameters; the method of adaptive normalization processing is as follows: Where s(n) is the interval of quantization operations, and n is the quantization bit width; in the point cloud feature distribution map, different channels have non-zero mean values, and the channel mean values ​​are not equal between different block sets; therefore, before performing further quantization operations, it is necessary to use the above calculation method to perform adaptive normalization processing based on the quantization bit width on the channel features of each channel; the adaptive normalization processing based on the quantization bit width on each channel feature also includes: after the activation function layer, determining the information type of the channel feature; if the information type of the channel feature is XYZ position information, the corresponding quantization bit width is 4 bits; if the information type of the channel feature is RGB color information or I intensity information, the corresponding quantization bit width is 8 bits; The training module is used to perform quantitative training on the model based on the channel feature parameters using backpropagation to obtain quantized data. This includes: statistically analyzing the mean and variance obtained from forward propagation before performing backpropagation of the neural network, summing the output of the convolution along the feature channel dimension, and performing precision conversion again using the saved mean and variance during the summation process to ensure that the network is differentiable.

6. A model quantization system for point cloud data, characterized in that, The system includes a memory and a processor. The memory stores a model quantization program for point cloud data, and the processor runs the model quantization program for the point cloud data to enable the model quantization system for the point cloud data to perform the model quantization method for the point cloud data as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a model quantization program for point cloud data, which, when executed by a processor, implements the model quantization method for point cloud data as described in any one of claims 1-4.

Images