Asymmetric double-branch DGB-D semantic segmentation method and device
The asymmetric dual-branch DGB-D semantic segmentation method combines convnext and Segformer networks with attention-based fusion to address format differences and inefficiencies in RGB-D image processing, enhancing segmentation accuracy and efficiency.
Patent Information
- Application Number
- CN202411126069.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2024-12-13
AI Technical Summary
Existing semantic segmentation algorithms for RGB-D images either ignore the format differences between RGB and Depth images or use symmetrical network structures that are not time-efficient, especially in poor lighting conditions or complex scenes.
An asymmetric dual-branch DGB-D semantic segmentation method using a convnext neural network for the RGB branch and Segformer transformer for the Depth branch, with an attention-based feature fusion algorithm to combine the branches, employing an encoder-decoder structure for pixel-level segmentation.
The method achieves efficient and effective semantic segmentation by leveraging the strengths of both RGB and Depth images, improving performance in challenging conditions through reduced computational complexity and optimized network structures.
Smart Images
Figure CN119131384A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision technology, and in particular relates to an asymmetric dual-branch DGB-D semantic segmentation method and device. Background Art
[0002] Semantic segmentation in the field of computer vision has always been a research direction that has attracted much attention, providing environmental perception for subsequent scene understanding, mapping and positioning tasks. Due to the limitations of hardware devices or data collection methods, the information captured by a single type of image is not enough to fully represent the imaging scene. Compared with a single image, the use of GBD images and Depth images can obtain both the color and edge perception capabilities of RGB cameras and the distance perception capabilities of depth cameras, which have better scene representation and visual perception capabilities, and stronger anti-interference capabilities.
[0003] At present, the mainstream semantic segmentation algorithms are all based on DGB images. The network structure for processing a single DGB image is simple and efficient, and there are already many mature lightweight networks and models. However, in poor lighting conditions or some complex scenes, it is necessary to use the assistance of depth images, which is called DGB-D semantic segmentation. There are two main approaches to the existing research on RGB-D semantic segmentation. The first is to directly splice the RGB and depth images in the channel after processing and then use them as the input of the transformer, but this method ignores the huge differences in the feature formats of the two images; the other is to use two identical branches to process RGB and Depth images, adopt a symmetrical network structure and fuse the two branches, but this method uses two branches and is not time-efficient. Summary of the invention
[0004] The object of the present invention is to provide an asymmetric dual-branch DGB-D semantic segmentation method and device to solve the above-mentioned technical problems.
[0005] In order to solve the above technical problems, the specific technical solutions of an asymmetric dual-branch DGB-D semantic segmentation method and device of the present invention are as follows:
[0006] An asymmetric dual-branch DGB-D semantic segmentation method comprises the following steps:
[0007] Step 1: Use the convnext neural network to build the RGB main branch in the encoder;
[0008] Step 2: Use the Segformer transformer to build the Depth sub-branch in the encoder;
[0009] Step 3: Use the feature fusion method based on the attention mechanism to fuse the dual-branch feature maps and update the main branch;
[0010] Step 4: The decoder obtains the final semantic segmentation result.
[0011] Furthermore, the step 1 comprises the following steps:
[0012] Step 1.1: Construct the convnext neural network basic unit module;
[0013] Step 1.2: Based on the basic modules of the neural network described in step 1.1, construct 4 RGB main branch backbone network modules.
[0014] Furthermore, the step 1.1 includes the following steps:
[0015] The basic module of the network adopts a residual network structure similar to Resnet and draws on the design ideas of SwinTransformer: for the input feature map x, a convolution kernel of 7 is first performed.
[0016] ×7 depth convolution, normalized using Layer Normalization, and then two layers of point convolution to expand the number of channels by four times, using GELU as the activation function, and then reducing the number of channels to one-fourth, applying DropPath to the processed data and adding it to the original input.
[0017] Furthermore, the step 1.2 includes the following steps:
[0018] For the downsampling of the first layer input, a convolution with a kernel of 4×4 and a stride of 4 is used. For the downsampling of the other three layers, a convolution with a kernel of 2×2 and a stride of 2 is used. The number of channels of the special graph is adjusted to [96, 192, 384, 768] through the downsampling convolution. After downsampling, the convolution layers are stacked separately.
[0019] [3,3,9,3] The basic modules constructed in step 1.1 are used to obtain a four-layer backbone network, and each layer is used as an independent block in subsequent steps.
[0020] Furthermore, the step 2 includes the following specific steps:
[0021] Step 2.1: Build the basic module of transformer;
[0022] Step 2.2: Based on the transformer basic module described in step 2.1, construct 4 Depth sub-branch trunk network modules.
[0023] Furthermore, the step 2.1 includes the following specific steps:
[0024] The basic module of the transformer consists of two parts: the efficient self-attention module Self-Atten and the mixed feedforward network module Mix-FNN.
[0025] In the original multi-head self-attention process, each head Q, K, V has the same dimension N×C, where N=H×W is the length of the sequence, and the self-attention estimate is:
[0026]
[0027] The computational complexity of this process is O(N 2 ), introduce a sequence reduction process and use the reduction ratio R to reduce the length of the sequence:
[0028]
[0029]
[0030] Where K is the sequence to be reduced, Reshape means reshaping the shape of K, Linear(C in ,C out ) refers to a linear layer with a C in dimensional tensor as input and generate a C out dimensional tensor as output, so the dimension of the new K is N / R×C, so the complexity of the self-attention mechanism is reduced from O(N 2 ) is reduced to O(N 2 / R), R is set to [64, 16, 4, 1] from the first layer to the fourth layer of the backbone network. Furthermore, the output of the self-attention module is:
[0031] x out =x in +Drop(Attention(Norm(x in )))
[0032] The operation of Mix-FFN is shown in the following formula:
[0033] x out =MLP(GELU(Conv 3×3 (MLP(x in ))))+x in
[0034] Among them, x in are the features from the self-attention module.
[0035] Furthermore, the step 2.2 includes the following specific steps:
[0036] The basic process of each module is to first divide the image patch, then stack the basic modules, and finally normalize and reshape the feature map. In the image patch-embed module, the image size and number of channels are adjusted through 2d strided convolution. In the first layer, a convolution with a kernel of 7×7 and a stride of 4 is used, and in the next three layers, a convolution with a kernel of 3×3 and a stride of 2 is used. H and W are flattened into one dimension and input into subsequent modules. In the basic module stacking operation, each layer stacks 2 basic modules constructed in step 2.1. After normalization, the feature map is reshaped according to the previously adjusted H and W as the output of the backbone network module.
[0037] Furthermore, the step 3 comprises the following steps:
[0038] Step 3.1: Construct a feature fusion module based on the attention mechanism;
[0039] The fusion method includes two parts: local attention-guided channel feature selection and cross-modal attention-guided feature correlation embedding. First, local attention-guided channel feature selection is performed on the input RGB and Depth images to obtain Fused1. Then, the Fused1 obtained in the previous step and the two original feature maps are used to perform cross-modal attention-guided feature correlation embedding to obtain the result Fused2. The results of the above two steps are directly added to obtain the final result.
[0040] Step 3.2: Construct the fusion network framework, only fuse the 2nd, 3rd and 4th layers to get the results Fused2, Fused3 and Fused4, then modify the RGB main network based on the fusion results, add Fused2 to the output of the second layer as the input of the third layer, and add Fused3 to the output of the third layer as the input of the fourth layer.
[0041] Furthermore, step 4 includes the following steps:
[0042] Using the output RGB1 of the first layer of the RGB main branch and the fusion results Fused2, Fused3, and Fused4 obtained in step 3.2 as the four-layer input of the decoder, an MLP decoder is proposed: first, the 4-level features Fi from the encoder pass through an MLP layer to unify the channel dimension; then, in the second step, the features are upsampled to 1 / 4 of the original image size and concatenated together; third, an MLP layer is used to fuse the concatenated features f; finally, another MLP layer uses the fused features to obtain H / 4×W / 4×N cls Resolution prediction segmentation result M, where N cls is the number of categories, and the decoder is expressed as:
[0043]
[0044] M=Linear(C,N cls )(F)
[0045] The predicted results are then upsampled four times to obtain the final semantic segmentation result.
[0046] The present invention also discloses an asymmetric dual-branch DGB-D semantic segmentation device, comprising a memory and one or more processors, wherein the memory stores executable code, and is characterized in that when the one or more processors execute the executable code, they are used to implement the steps of the asymmetric dual-branch DGB-D semantic segmentation method.
[0047] The asymmetric dual-branch DGB-D semantic segmentation method and device of the present invention have the following advantages:
[0048] The invention discloses an asymmetric dual-branch DGB-D semantic segmentation method, which adopts an encoder-decoder structure. The RGB branch is used as the main branch in the encoder, and the network structure is more complex, and the parameters and calculation amount are larger; the Depth branch is used as the secondary branch, and the network structure is simpler, and the parameters and calculation amount are smaller; the two branches are fused using an attention-based feature fusion algorithm, and the main branch is updated according to the fusion result. Finally, pixel-level semantic segmentation is achieved according to the result obtained by the MLP decoder. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 The network structure diagram proposed by the present invention;
[0050] Figure 2 This is a structural diagram of the RGB-D fusion module used in the present invention;
[0051] Figure 3 It is a schematic diagram of the physical structure of the device according to the embodiment of the present invention. DETAILED DESCRIPTION
[0052] In order to better understand the purpose, structure and function of the present invention, the following is a further detailed description of an asymmetric dual-branch DGB-D semantic segmentation method and device of the present invention in conjunction with the accompanying drawings.
[0053] like Figure 1 As shown, an asymmetric dual-branch DGB-D semantic segmentation method of the present invention comprises the following steps:
[0054] Step 1: Use the convnext neural network to build the RGB main branch in the encoder.
[0055] The process of building the RGB main branch in the encoder using the convnext neural network includes the following steps:
[0056] Step 1.1: Construct the basic unit module of the convnext neural network. The basic module of the network adopts a residual network structure similar to Resnet and draws on the design ideas of Swin Transformer. For the input feature map x, first perform a deep convolution with a convolution kernel of 7×7 and use Layer Normalization for normalization. Then perform two layers of point convolution to expand the number of channels by four times, use GELU (Gaussian Error Linear Unit) as the activation function, and then reduce the number of channels to one-fourth. Apply DropPath to the processed data and add it to the original input.
[0057] Step 1.2: According to the basic modules of the neural network described in step 1.1, construct 4 RGB main branch backbone network modules. Since the basic modules in the above steps do not change the size and number of channels of the feature map, the backbone network of each layer should include two parts: a downsampling module and a stack of basic modules. For the downsampling of the first layer input, a convolution with a kernel of 4×4 and a stride of 4 is used. For the downsampling of the other three layers, a convolution with a kernel of 2×2 and a stride of 2 is used; and the number of channels of the feature map is adjusted to [96, 192, 384, 768] through the downsampling convolution. After downsampling, stack [3, 3, 9, 3] basic modules constructed in step 1.1 respectively to obtain a four-layer backbone network. Each layer is used as an independent block in the subsequent steps.
[0058] Step 2: Use the Segformer transformer to build the Depth side branch in the encoder.
[0059] Step 2.1: Build the basic module of transformer: It consists of two parts: efficient self-attention module (Self-Atten) and mixed feedforward network module (Mix-FNN).
[0060] In the original multi-head self-attention process, each head Q, K, V has the same dimension N×C, where N=H×W is the length of the sequence, and the self-attention estimate is:
[0061]
[0062] The computational complexity of this process is O(N 2 ), which is very large for an image, so a sequence reduction process is introduced, using the reduction ratio R to reduce the length of the sequence:
[0063]
[0064]
[0065] Where K is the sequence to be reduced, Reshape means reshaping the shape of K, Linear(C in ,C out ) refers to a linear layer with a C in dimensional tensor as input and generate a C out dimensional tensor as output. Therefore, the dimension of the new K is N / R×C, so the complexity of the self-attention mechanism is reduced from O(N 2 ) is reduced to O(N 2 / R). Set R to [64, 16, 4, 1] from the first to the fourth layers of the backbone network. Furthermore, the output of the self-attention module is:
[0066] x out =x in +Drop(Attention(Norm(x in )))
[0067] The mixed feedforward network is designed to introduce position information in a data-driven way without the need for position encoding. Mix-FFN achieves this by combining a 3×3 convolutional layer with a multilayer perceptron (MLP) and using deep convolution to reduce the number of parameters and improve efficiency. Specifically, the operation of Mix-FFN can be expressed as the following formula:
[0068] x out =MLP(GELU(Conv 3×3 (MLP(x in ))))+x in
[0069] Among them, x in are the features from the self-attention module.
[0070] Step 2.2: Based on the transformer basic module described in step 2.1, construct 4 depth sub-branch backbone network modules. The basic process of each module is to first divide the image patch, then stack the basic modules, and finally normalize and reshape the feature map.
[0071] In the image Patch-Embed module, the image size and number of channels are adjusted mainly through 2d strided convolution. In the first layer, convolution with a kernel of 7×7 and a stride of 4 is used, and in the next three layers, convolution with a kernel of 3×3 and a stride of 2 is used. H and W are flattened into one dimension and input into the subsequent modules. In the basic module stacking operation, each layer stacks 2 basic modules built in step 2.1. After normalization, the feature map is reshaped according to the previously adjusted H and W as the output of the backbone network module.
[0072] Step 3: Use the feature fusion method based on the attention mechanism to fuse the dual-branch feature maps and update the main branch.
[0073] Step 3.1: Construct a feature fusion module based on the attention mechanism. The method proposed in the present invention does not impose any specific restrictions on the selection and construction of feature modules, but it should be noted that the two branches to be fused have different numbers of channels and cannot be added. Here is an example of a fusion module, such as Figure 2 As shown in the figure, the proposed fusion method mainly includes two parts: local attention-guided channel feature selection and cross-modal attention-guided feature correlation embedding. First, local attention-guided channel feature selection is performed on the input RGB and Depth images to obtain Fused1. Then, the Fused1 obtained in the previous step and the two original feature maps are used to perform cross-modal attention-guided feature correlation embedding to obtain the result Fused2. The results of the above two steps are directly added to get the final result.
[0074] Step 3.2: Construct the fusion network framework. Steps 1 and 2 result in a four-layer dual-branch backbone network. Since the feature map processed only by the first layer is large, the fusion calculation consumes too much memory and computing resources, so only the second, third, and fourth layers are actually fused to obtain the results Fused2, Fused3, and Fused4. Then, the RGB main network is corrected based on the fusion results. Specifically, the output of Fused2 and the second layer is added as the input of the third layer, and the output of Fused3 and the third layer is added as the input of the fourth layer.
[0075] Step 4: The decoder obtains the final semantic segmentation result.
[0076] Use the output RGB1 of the first layer of the RGB main branch and the fusion results Fused2, Fused3, and Fused4 obtained in step 3.2 as the four-layer input of the decoder.
[0077] The proposed MLP decoder consists of four main steps. First, the 4-level features Fi from the encoder are passed through an MLP layer to unify the channel dimension. Then, in the second step, the features are upsampled to 1 / 4 of the original image size and concatenated together. Third, an MLP layer is used to fuse the concatenated features f. Finally, another MLP layer uses the fused features to generate H / 4×W / 4×N cls Resolution prediction segmentation result M, where N cls is the number of categories. This allows us to formulate the decoder as:
[0078]
[0079] M=Linear(C,N cls )(F)
[0080] The prediction results obtained in the above steps are then upsampled four times to obtain the final semantic segmentation result.
[0081] Corresponding to the aforementioned embodiment of an asymmetric dual-branch DGB-D semantic segmentation method, the present invention also provides an asymmetric dual-branch DGB-D semantic segmentation device embodiment, as shown in the attached Figure 3 As shown. It includes a memory and one or more processors, and the memory contains all executable codes, training models and training data sets. When the processor executes the executable code, an asymmetric dual-branch DGB-D semantic segmentation method is implemented.
[0082] An embodiment of an asymmetric dual-branch DGB-D semantic segmentation method provided by the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located reading the corresponding computer program instructions in the non-volatile memory into the internal memory for execution. From a hardware perspective, if Figure 3 As shown, the present invention provides a hardware structure diagram of an asymmetric dual-branch DGB-D semantic segmentation device in any device with data processing capability, except Figure 3 In addition to the processor, memory, network interface, multi-pipeline image acquisition device, and fixed memory shown, any device with data processing capability in which the device in the embodiment is located may also include other hardware according to the actual function of the device with data processing capability, which will not be described in detail. The implementation process of the functions and effects of each unit in the above device is specifically detailed in the implementation process of the corresponding steps in the above method, which will not be described in detail here.
[0083] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.
[0084] An embodiment of the present invention also provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, an asymmetric dual-branch DGB-D semantic segmentation method in the above embodiment is implemented. The computer-readable storage medium can be an internal storage unit of any device with data processing capabilities described in any of the above embodiments, such as a hard disk or a memory. The computer-readable storage medium can also be an external storage device of any device with data processing capabilities, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), an SD card, a flash card (Flash Card), a graphics card (Graphics Card), etc. equipped on the device. Further, the computer-readable storage medium can also include both an internal storage unit and an external storage device of any device with data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store data that has been output or is to be output.
[0085] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.
Claims
1. An asymmetric dual-branch DGB-D semantic segmentation method, characterized in that: The steps include: Step 1: Use the convnext neural network to build the RGB main branch in the encoder; Step 2: Use the Segformer transformer to build the Depth sub-branch in the encoder; Step 3: Use the feature fusion method based on the attention mechanism to fuse the dual-branch feature maps and update the main branch; Step 4: The decoder obtains the final semantic segmentation result.
2. The asymmetric dual-branch DGB-D semantic segmentation method according to claim 1, characterized in that: The step 1 comprises the following steps: Step 1.1: Construct the convnext neural network basic unit module; Step 1.2: Based on the basic modules of the neural network described in step 1.1, construct 4 RGB main branch backbone network modules.
3. The asymmetric dual-branch DGB-D semantic segmentation method according to claim 2, characterized in that: The step 1.1 comprises the following steps: The basic module of the network adopts a residual network structure similar to Resnet and draws on the design idea of Swin Transformer: for the input feature map x, a deep convolution with a convolution kernel of 7×7 is first performed, normalized using LayerNormalization, and then two layers of point convolution are performed to expand the number of channels by four times, and GELU is used as the activation function, and then the number of channels is reduced to one-fourth. DropPath is applied to the processed data and added to the original input.
4. The asymmetric dual-branch DGB-D semantic segmentation method according to claim 3, characterized in that: The step 1.2 comprises the following steps: For the downsampling of the first layer input, a convolution with a kernel of 4×4 and a stride of 4 is used. For the downsampling of the other three layers, a convolution with a kernel of 2×2 and a stride of 2 is used. The number of channels of the special graph is adjusted to [96, 192, 384, 768] through the downsampling convolution. After downsampling, the basic modules constructed in step 1.1 are stacked [3, 3, 9, 3] respectively to obtain a four-layer backbone network. Each layer is applied as an independent block in the subsequent steps.
5. The asymmetric dual-branch DGB-D semantic segmentation method according to claim 1, characterized in that: The step 2 comprises the following specific steps: Step 2.1: Build the basic module of transformer; Step 2.2: Based on the transformer basic module described in step 2.1, construct 4 Depth sub-branch trunk network modules.
6. The asymmetric dual-branch DGB-D semantic segmentation method according to claim 5, characterized in that: The step 2.1 includes the following specific steps: The basic module of the transformer consists of two parts: the efficient self-attention module Self-Atten and the mixed feedforward network module Mix-FNN. In the original multi-head self-attention process, each head Q, K, V has the same dimension N×C, where N=H×W is the length of the sequence, and the self-attention estimate is: The computational complexity of this process is O(N 2 ), introduce a sequence reduction process and use the reduction ratio R to reduce the length of the sequence: Where K is the sequence to be reduced, Reshape means reshaping the shape of K, Linear(C in ,C out ) refers to a linear layer with a C in dimensional tensor as input and generate a C out dimensional tensor as output, so the dimension of the new K is N / R×C, so the complexity of the self-attention mechanism is reduced from O(N 2 ) is reduced to O(N 2 / R), R is set to [64, 16, 4, 1] from the first layer to the fourth layer of the backbone network. Furthermore, the output of the self-attention module is: x out =x in +Drop(Attention(Norm(x in ))) The operation of Mix-FFN is shown in the following formula: x out =MLP(GELU(Conv 3×3 (MLP(x in ))))+x in Among them, x in are the features from the self-attention module.
7. The asymmetric dual-branch DGB-D semantic segmentation method according to claim 5, characterized in that: The step 2.2 includes the following specific steps: The basic process of each module is to first divide the image patch, then stack the basic modules, and finally normalize and reshape the feature map. In the image patch-embed module, the image size and number of channels are adjusted through 2d strided convolution. In the first layer, a convolution with a kernel of 7×7 and a stride of 4 is used, and in the next three layers, a convolution with a kernel of 3×3 and a stride of 2 is used. H and W are flattened into one dimension and input into subsequent modules. In the basic module stacking operation, each layer stacks 2 basic modules constructed in step 2.
1. After normalization, the feature map is reshaped according to the previously adjusted H and W as the output of the backbone network module.
8. The asymmetric dual-branch DGB-D semantic segmentation method according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1: Construct a feature fusion module based on the attention mechanism; The fusion method includes two parts: local attention-guided channel feature selection and cross-modal attention-guided feature correlation embedding. First, local attention-guided channel feature selection is performed on the input RGB and Depth images to obtain Fused1. Then, the Fused1 obtained in the previous step and the two original feature maps are used to perform cross-modal attention-guided feature correlation embedding to obtain the result Fused2. The results of the above two steps are directly added to obtain the final result. Step 3.2: The fusion network framework is constructed. Only the 2nd, 3rd and 4th layers are fused to obtain the results Fused2, Fused3 and Fused4. Then, the RGB main network is modified based on the fusion results. The output of Fused2 and the second layer is added as the input of the third layer, and the output of Fused3 and the third layer is added as the input of the fourth layer.
9. The asymmetric dual-branch DGB-D semantic segmentation method according to claim 8, characterized in that: The step 4 comprises the following steps: Using the output RGB1 of the first layer of the RGB main branch and the fusion results Fused2, Fused3, and Fused4 obtained in step 3.2 as the four-layer input of the decoder, an MLP decoder is proposed: first, the 4-level features Fi from the encoder pass through an MLP layer to unify the channel dimension; then, in the second step, the features are upsampled to 1 / 4 of the original image size and concatenated together; third, an MLP layer is used to fuse the concatenated features f; finally, another MLP layer uses the fused features to obtain H / 4×W / 4×N cls Resolution prediction segmentation result M, where N cls is the number of categories, and the decoder is expressed as: M=Linear(C,N cls )(F) The predicted results are then upsampled four times to obtain the final semantic segmentation result.
10. An asymmetric dual-branch DGB-D semantic segmentation device, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the one or more processors execute the executable code, they are used to implement the steps of an asymmetric dual-branch DGB-D semantic segmentation method according to any one of claims 1 to 9.