Multi-task intensive prediction method based on mixed single-rank experts

Abstract

The invention discloses a multi-task intensive prediction method based on mixed single-rank experts. The method comprises the following steps: collecting and arranging a multi-task intensive data set; dividing the data set into a training set, a verification set and a test set, and performing preprocessing; constructing a multi-task intensive prediction model; training the model by using a training set, and regularly evaluating the performance of the model in the training process by using a verification set; carrying out loss calculation and parameter optimization on the model in the training process; and performing final performance evaluation on the trained model by using the test set, and applying the trained model to a to-be-reasoned image to obtain an output result of the image under each dense prediction task. According to the method, a model comprising a shared feature extraction module, a task feature enhancement module and a task prediction module is constructed, and a mixed single-rank expert group is adopted to dynamically enhance task features in space and channel dimensions, so that a unified learning framework for performing multi-task intensive prediction on images is realized.

Description

Technical Field

[0001] This invention relates to the field of computer vision multi-task learning technology, and in particular to a multi-task dense prediction method based on hybrid single-rank experts. Background Technology

[0002] In recent years, large-scale pre-trained models have made significant progress in the field of computer vision. These models, by utilizing massive amounts of data for general feature learning, possess strong generalization capabilities. However, in practical applications, visual models often need to be trained specifically for particular tasks to achieve better predictive performance. For example, in intensive prediction tasks such as semantic segmentation, saliency estimation, and edge detection, single-task retraining is typically required on top of the pre-trained model. This single-task fine-tuning approach not only increases the cost of training and deployment but also makes it difficult to fully utilize the potential correlations between different tasks, thus limiting the model's efficiency and generalization ability in multi-task scenarios.

[0003] To address this issue, researchers have gradually introduced multi-task learning methods. This approach enables models to simultaneously handle multiple visually intensive prediction tasks within a unified framework, achieving efficient joint training and inference. Early methods primarily achieved collaborative learning between tasks by designing sophisticated network structures. For example, at the encoder level, each task shares a feature extraction backbone to jointly acquire general representations; at the decoder level, task-specific decoding modules are introduced, allowing each task to learn more discriminative features in independent decoding paths. While these methods can indeed improve multi-task prediction performance to some extent, their network structures are statically pre-defined, making it difficult to adaptively construct task-specific features when task requirements change, and also difficult to effectively capture the global relationships between tasks.

[0004] Recently, to further improve the performance of multi-task learning methods, some studies have introduced hybrid expert techniques, such as MoE (Mixture of Experts). This method adaptively selects computational paths among different experts through a dynamic routing mechanism to model the differences between tasks and improve the expressive power of task-specific features. While these methods have achieved some success at the encoder level, research on the decoder level is still insufficient. As the number of experts increases, the parameter size and computational cost of the model also rise significantly, thus limiting the scalability of these methods in practical applications. To address this issue, some studies have proposed hybrid low-rank expert methods for the decoder, controlling the parameter size of individual experts by constraining the rank within each expert. However, this method still has some problems. For example, low-rank experts compress features into a low-dimensional subspace, which may lead to dimensional redundancy among different experts, limiting the feature representation ability of a single expert. Simultaneously, the globally shared hybrid structure lacks the ability to simultaneously model general knowledge across tasks and specific knowledge within tasks, making it difficult for the overall expert structure to form a feature representation that better balances sharing and task specificity. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology. To achieve the above objective, a multi-task dense prediction method based on hybrid single-rank experts is adopted to solve the problems mentioned in the background technology.

[0006] A multi-task dense prediction method based on hybrid single-rank experts includes the following steps: Step S1: Collect and organize multi-task intensive datasets, where each image in the dataset contains multiple types of dense annotation information; Step S2: Divide the dataset into training set, validation set and test set, and preprocess the images in each set to make them conform to the model input format; Step S3: Construct a multi-task dense prediction model based on hybrid single-rank experts; the model includes a shared feature extraction module, a task feature enhancement module, and a task prediction module; the shared feature extraction module is used to extract multi-level shared general features from the input image; the task feature enhancement module is used to enhance the structural and semantic information of the shared general features sequentially in the spatial and channel dimensions, wherein each processing unit consists of a hybrid single-rank expert group that is shared by the task and specific to the task, and the hybrid single-rank expert group includes a routing module and multiple expert modules with rank 1; the task prediction module is used to receive the enhanced features and output the final prediction results for each task; Step S4: Train the model using the training set, and periodically evaluate the model performance using the validation set during the training process; during the training process, a multi-task loss joint optimization strategy is used to calculate the loss and optimize the parameters of the model. Step S5: Use the test set to perform a final performance evaluation on the trained model, and use the trained model on the image to be inferred to obtain the output results of the image under each dense prediction task.

[0007] As a further aspect of the present invention: the shared feature extraction module is based on the Transformer architecture and specifically performs the following operations: The input image is segmented and positional encoding information is added. This segment is then fed into multiple Transformer blocks for multi-head self-attention computation. The outputs of four specific Transformer blocks are extracted from low to high resolution, and the output features of adjacent blocks are aggregated to obtain common features shared by the three tasks. The aggregation process is represented as follows:

[0008] in, express Convolution operation, This indicates a channel splicing operation. express The feature index and the corresponding processing module index.

[0009] As a further aspect of the present invention: the task feature enhancement module includes five enhancement modules for any task, wherein the first three modules process the corresponding shared general features, and the last two modules are used to further process the fused task enhancement features, specifically as follows:

[0010]

[0011] =

[0012]

[0013]

[0014] =

[0015] in, Indicates the task index. Used to represent feature indexes and corresponding processing module indexes. This indicates the output characteristics of the enhancement module. Indicates the weighting coefficient. This represents the output after element-wise feature aggregation. This represents the output characteristics after processing by all enhancement modules.

[0016] As a further aspect of the present invention: each feature enhancement module is composed of a task-adaptive convolution, a spatial processing unit, and a channel processing unit connected in sequence, and the processing procedure is represented as follows:

[0017]

[0018]

[0019]

[0020] in, As input to the feature enhancement module, This is the output of the feature enhancement module.

[0021] As a further aspect of the present invention: the spatial processing unit consists of a hybrid single-rank expert group (SAG) combining task sharing and task-specific methods, along with convolutional operations; the hybrid single-rank expert group includes a routing module and... Each expert module has a rank of 1; each expert consists of... and convolution kernel Composition; the processing procedure of the space processing unit is specifically represented as follows:

[0022]

[0023]

[0024]

[0025]

[0026]

[0027]

[0028] in, , , These represent the channel dimension, height, and width of the feature map, respectively. Indicates the channel index. This indicates a global average pooling operation. Indicates selecting from largest to smallest. k The index of each value This represents the selected set of expert indexes. Represents the temperature term. This indicates a custom convolution operation. Represents the convolution kernel parameters. This indicates the processing results of a hybrid single-rank expert group that shares tasks. This indicates the processing result of a task-specific hybrid single-rank expert group.

[0029] As a further aspect of the present invention: the channel processing unit consists of a hybrid single-rank expert group comprising task-sharing and task-specific methods, and convolutional operations; each expert is composed of... and parameter vector Composition; the processing procedure of the channel processing unit is specifically represented as follows:

[0030]

[0031]

[0032]

[0033]

[0034]

[0035]

[0036] in, Indicates batch size, This indicates a dimensional change operation. Indicates the first Each token characteristic, This represents the matrix multiplication operation.

[0037] As a further aspect of the present invention: for the task The prediction module operates as follows:

[0038] in, Indicates an upsampling operation. For the model to the task The dense prediction results.

[0039] As a further aspect of the present invention: the multi-task loss joint optimization strategy adopts a weighted multi-task joint loss as the overall optimization objective, and the calculation of the joint loss is expressed as follows:

[0040] in, Indicates the task index. Indicates the total number of tasks. Indicates the first The weight of the loss term for each task, The overall optimization objective of the model.

[0041] As a further aspect of the present invention: the loss function is selected according to the task type, including: The semantic segmentation task employs the cross-entropy loss function; The edge detection task employs a balanced cross-entropy loss function. The surface normal estimation task uses the L1 loss function.

[0042] As a further aspect of the present invention: the performance evaluation adopts evaluation metrics corresponding to the task, including: using the average intersection-union ratio to evaluate the semantic segmentation task, using the root mean square error to evaluate the surface normal estimation task, and using the optimal dataset scale F-value to evaluate the edge detection task.

[0043] Compared with the prior art, the present invention has the following technical advantages: Using the aforementioned technical solution, a dedicated model was constructed comprising a shared feature extraction module, a task feature enhancement module, and a task prediction module. Its core lies in the introduction of a hybrid single-rank expert group—combining task-shared and task-specific methods—into the task feature enhancement module. This expert group dynamically selects and combines multiple rank-1 experts through a routing module, adaptively enhancing the shared features in both spatial and channel dimensions. Combined with a multi-task loss joint optimization strategy, this method can simultaneously output multiple dense prediction results from a single image.

[0044] Hybrid single-rank expert design significantly improves the parameter efficiency of the model. While maintaining a compact model size, it achieves efficient and specialized enhancement of multi-task features through low-rank decomposition and dynamic routing mechanisms. Secondly, this method effectively alleviates the "negative transfer" problem in multi-task learning. Task-specific experts can capture unique patterns, while task-sharing experts extract general knowledge, thereby synergistically improving the performance of each task. Finally, the unified feature enhancement and joint optimization framework enhances the model's generalization ability and deployment flexibility, enabling it to achieve accurate and efficient multi-task joint inference in complex visual scenes. Attached Figure Description

[0045] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the multi-task intensive prediction method of the present invention.

[0046] Figure 2 This is a schematic diagram of the overall architecture of the model of the present invention.

[0047] Figure 3 This is a schematic diagram of the task feature enhancement module structure of the present invention.

[0048] Figure 4 This is a schematic diagram of the spatial enhancement processing unit structure of the present invention.

[0049] Figure 5 This is a schematic diagram of the channel enhancement processing unit structure of the present invention.

[0050] Figure 6 This is a schematic diagram of the hybrid single-rank expert structure of the present invention.

[0051] Figure 7 This is a schematic diagram of the task prediction module structure of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Please refer to Figure 1 In this embodiment of the invention, a multi-task dense prediction method based on hybrid single-rank experts includes the following steps: Step S1: Collect and organize multi-task intensive datasets, where each image in the dataset contains multiple types of dense annotation information; In this embodiment, each image in the dataset has different types of dense annotation information, such as semantic class information and edge class information of each pixel; Step S2: Divide the dataset into training set, validation set and test set, and preprocess the images in each set to make them conform to the model input format, that is, perform interpolation and other operations to make them conform to the model input format requirements. like Figure 2 As shown in the figure, the diagram is a schematic diagram of the overall architecture of the model.

[0054] Step S3: Construct a multi-task dense prediction model based on hybrid single-rank experts; the model includes a shared feature extraction module, a task feature enhancement module, and a task prediction module; the shared feature extraction module is used to extract multi-level shared general features from the input image; the task feature enhancement module is used to enhance the structural and semantic information of the shared general features sequentially in the spatial and channel dimensions, wherein each processing unit consists of a hybrid single-rank expert group that is shared by the task and specific to the task, and the hybrid single-rank expert group includes a routing module and multiple expert modules with rank 1; the task prediction module is used to receive the enhanced features and output the final prediction results for each task; In this embodiment, the shared feature extraction module is based on the Transformer architecture. The shared feature extraction module is used to extract multi-level task-shared features from the input image, specifically performing the following operations: Input image The process involves performing a patch operation and adding positional encoding information, then inputting the data into Transformer Blocks, where multiple Transformer blocks undergo multi-head self-attention computation. After computation, the outputs of four specific Transformer blocks are extracted from lowest to highest resolution. The output features of adjacent blocks are aggregated to obtain common features shared by the three tasks. The polymerization process is represented as:

[0055] in, express Convolution operation, This indicates a channel splicing operation. express The feature index and the corresponding processing module index.

[0056] like Figure 3 As shown in the figure, this is a schematic diagram of the task feature enhancement module structure.

[0057] In this embodiment, the task feature enhancement module is used to enhance the structural and semantic information of the shared features sequentially in the spatial and channel dimensions. Task feature enhancement module for any task Each contains five enhancement modules. The first three modules process respectively The corresponding shared general features are used in the first two modules, which are used to further process the fused task enhancement features. The specific process is as follows:

[0058]

[0059] =

[0060]

[0061]

[0062] =

[0063] in, Indicates the task index. Used to represent feature indexes and corresponding processing module indexes. This indicates the output characteristics of the enhancement module. Indicates the weighting coefficient. This represents the output after element-wise feature aggregation. This represents the output characteristics after processing by all enhancement modules.

[0064] like Figure 4 As shown in the figure, it is a schematic diagram of the spatial enhancement processing unit structure.

[0065] like Figure 5 As shown in the figure, this is a schematic diagram of the channel enhancement processing unit structure.

[0066] In this embodiment, each feature enhancement module is a task-adaptive convolution. Space processing unit and channel processing unit The components are connected sequentially, and the processing procedure is represented as follows:

[0067]

[0068]

[0069]

[0070] in, As input to the feature enhancement module, This is the output of the feature enhancement module.

[0071] In this embodiment, the spatial processing unit consists of a hybrid single-rank expert group (SQA) combining task-shared and task-specific methods, along with convolutional operations; the hybrid single-rank expert group includes a routing module. and Expert modules with rank 1 For the space processing unit, each expert is... and convolution kernel Composition; the processing procedure of the space processing unit is specifically represented as follows:

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] in, , , These represent the channel dimension, height, and width of the feature map, respectively. Indicates the channel index. This indicates a global average pooling operation. Indicates selecting from largest to smallest. k The index of each value This represents the selected set of expert indexes. Represents the temperature term. This indicates a custom convolution operation. Represents the convolution kernel parameters. This indicates the processing results of a hybrid single-rank expert group that shares tasks. This indicates the processing result of a task-specific hybrid single-rank expert group.

[0079] like Figure 6 As shown, the diagram illustrates a hybrid single-rank expert structure.

[0080] In this embodiment, the channel processing unit consists of a hybrid single-rank expert group combining task-shared and task-specific methods, along with convolutional operations; each expert is composed of... and parameter vector Composition; the processing procedure of the channel processing unit is specifically represented as follows:

[0081]

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] in, Indicates batch size, This indicates a dimensional change operation. Indicates the first Each token characteristic, This represents the matrix multiplication operation.

[0088] like Figure 7 As shown in the figure, this is a schematic diagram of the task prediction module structure.

[0089] In this embodiment, for the task The prediction module operates as follows:

[0090] in, Indicates an upsampling operation. For the model to the task The dense prediction results.

[0091] Step S4: Train the model using the training set, and periodically evaluate the model performance using the validation set during the training process; during the training process, a multi-task loss joint optimization strategy is used to calculate the loss and optimize the parameters of the model. In this embodiment, the loss function is selected according to the task type, including: The semantic segmentation task employs the cross-entropy loss function; The edge detection task employs a balanced cross-entropy loss function. The surface normal estimation task uses the L1 loss function.

[0092] In this embodiment, a multi-task loss joint optimization strategy is adopted during training. The multi-task loss joint optimization strategy uses weighted multi-task joint loss as the overall optimization objective. The specific steps are as follows: During model training, a weighted multi-task joint loss is used as the overall optimization objective. For example, the standard cross-entropy loss function is used for semantic segmentation; the balanced cross-entropy loss function is used for edge detection to better handle the imbalance between positive and negative samples; and the L1 loss function is used for surface normal estimation to constrain continuous value prediction. By weighting and fusing the losses from these tasks, a unified optimization objective is formed, thereby achieving collaborative training and performance improvement across multiple tasks. The specific calculation steps are as follows:

[0093] in, This represents the weight of the loss term for the corresponding task. The overall optimization objective of the model.

[0094] The calculation of the joint loss is expressed as follows:

[0095] in, Indicates the task index. Indicates the total number of tasks. Indicates the first The weight of the loss term for each task, The overall optimization objective of the model.

[0096] Step S5: Use the test set to perform a final performance evaluation on the trained model, and use the trained model on the image to be inferred to obtain the output results of the image under each dense prediction task.

[0097] Specifically, a benchmark test is set, which includes the following tests: The trained model is then evaluated using test set data, as follows: In this embodiment, the performance evaluation adopts the evaluation index corresponding to the task, including: using the average intersection-union ratio to evaluate the semantic segmentation task, using the root mean square error to evaluate the surface normal estimation task, and using the optimal dataset scale F-value to evaluate the edge detection task.

[0098] Specifically, the evaluation indicators are implemented as follows, for example: (1) Mean Intersection over Union (mIoU): used to evaluate the prediction accuracy of semantic segmentation tasks and human body key part segmentation tasks; (2) Optimal dataset scale F-value (odsF): used to evaluate the performance of edge detection tasks; (3) Maximum F-value (maxF): Used to evaluate the performance of the significance estimation task; (4) Average angle error (mErr): used to evaluate the accuracy of the surface normal estimation task.

[0099] Example 2 Step 1: Organize the multi-task dataset. To ensure a fair comparison with other techniques, this embodiment uses the PASCAL-Context dataset. This dataset contains 10,103 images, each with annotations for semantic segmentation, human key part segmentation, saliency estimation, surface normal estimation, and edge detection.

[0100] Step 2: Split the dataset and perform preprocessing. For the PASCAL-Context dataset, the training set contains 4998 images, and the validation and test sets contain the same 5105 images. Interpolate the images in this dataset to achieve the desired resolution. .

[0101] Step 3: Construct a multi-task dense prediction model based on hybrid single-rank experts. For the shared feature extraction module, ViT-Large, pre-trained on ImageNet, was used. For each task feature enhancement module, four sets of hybrid single-rank experts were used. For low-level features, three sets of task-shared hybrid single-rank experts and one set of task-specific hybrid single-rank experts were used. For mid-level features, two sets of task-shared hybrid single-rank experts and two sets of task-specific hybrid single-rank experts were used. For high-level features, one set of task-shared hybrid single-rank experts and three sets of task-specific hybrid single-rank experts were used. Furthermore, for the first three task feature enhancement modules, eight single-rank experts were used as a group, with a TopK of 2. For the last two task feature enhancement modules, sixteen single-rank experts were used as a group, with a TopK of 4.

[0102] Step 4: Train the model using the training set, and use the validation set to periodically evaluate the model's current performance during training. A multi-task loss joint optimization strategy is employed during training to calculate loss and optimize parameters. The hyperparameter settings for the PASCAL-Context dataset are as follows: training batch size of 2, validation batch size of 6; 40,000 training iterations, validation interval of 2,000 iterations; Adam optimizer with a learning rate of 2e-5 and a weight decay coefficient of 1e-6; Poly learning rate scheduler; loss coefficients of 50.0 for edge detection, 10.0 for surface normal estimation, 1.0 for semantic segmentation, 2.0 for human key part segmentation, 5.0 for saliency estimation, and 0.0005 for expert load balancing.

[0103] Step 5: Perform final evaluation of the trained model. For the PASCAL-Context dataset, mIoU is used as the evaluation metric for semantic segmentation and human body key part segmentation; odsF is used as the performance metric for edge detection; maxF is used as the performance metric for saliency estimation; and mErr is used as the accuracy metric for surface normal estimation. For mIoU, MaxF, and odsF, higher values ​​are better; conversely, for mErr, lower values ​​are better.

[0104] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention. The scope of the invention is defined by the appended claims and their equivalents, all of which should be included within the scope of protection of the invention.

Claims

1. A multi-task dense prediction method based on hybrid single-rank experts, characterized in that, Includes the following steps: Step S1: Collect and organize multi-task intensive datasets, where each image in the dataset contains multiple types of dense annotation information; Step S2: Divide the dataset into training set, validation set and test set, and preprocess the images in each set to make them conform to the model input format; Step S3: Construct a multi-task dense prediction model based on hybrid single-rank experts; the model includes a shared feature extraction module, a task feature enhancement module, and a task prediction module; the shared feature extraction module is used to extract multi-level shared general features from the input image; the task feature enhancement module is used to enhance the structural and semantic information of the shared general features sequentially in the spatial and channel dimensions, wherein each processing unit consists of a hybrid single-rank expert group that is shared by the task and specific to the task, and the hybrid single-rank expert group includes a routing module and multiple expert modules with rank 1; the task prediction module is used to receive the enhanced features and output the final prediction results for each task; Step S4: Train the model using the training set, and periodically evaluate the model performance using the validation set during the training process; during the training process, a multi-task loss joint optimization strategy is used to calculate the loss and optimize the parameters of the model. Step S5: Use the test set to perform a final performance evaluation on the trained model, and use the trained model on the image to be inferred to obtain the output results of the image under each dense prediction task.

2. The multi-task dense prediction method based on hybrid single-rank experts according to claim 1, characterized in that, The shared feature extraction module is based on the Transformer architecture and performs the following operations: The input image is segmented and positional encoding information is added. This segment is then fed into multiple Transformer blocks for multi-head self-attention computation. The outputs of four specific Transformer blocks are extracted from low to high resolution, and the output features of adjacent blocks are aggregated to obtain common features shared by the three tasks. The aggregation process is represented as follows: in, express Convolution operation, This indicates a channel splicing operation. express The feature index and the corresponding processing module index.

3. The multi-task dense prediction method based on hybrid single-rank experts according to claim 2, characterized in that, The task feature enhancement module includes five enhancement modules for any task. The first three modules process the corresponding shared general features, and the last two modules are used to further process the fused task enhancement features. The specific process is as follows: = = in, Indicates the task index. Used to represent feature indexes and corresponding processing module indexes. This indicates the output characteristics of the enhancement module. Indicates the weighting coefficient. This represents the output after element-wise feature aggregation. This represents the output characteristics after processing by all enhancement modules.

4. The multi-task dense prediction method based on hybrid single-rank experts according to claim 3, characterized in that, Each feature enhancement module consists of a task-adaptive convolution, a spatial processing unit, and a channel processing unit connected sequentially. The processing procedure is as follows: in, As input to the feature enhancement module, This is the output of the feature enhancement module.

5. The multi-task dense prediction method based on hybrid single-rank experts according to claim 4, characterized in that, The spatial processing unit consists of a hybrid single-rank expert group (SQA) combining task-shared and task-specific methods, along with convolutional operations; the hybrid single-rank expert group includes a routing module and... Each expert module has a rank of 1; each expert consists of... and convolution kernel Composition; the processing procedure of the space processing unit is specifically represented as follows: in, , , These represent the channel dimension, height, and width of the feature map, respectively. Indicates the channel index. This indicates a global average pooling operation. Indicates selecting from largest to smallest. k The index of each value This represents the selected set of expert indexes. Represents the temperature term. This indicates a custom convolution operation. Represents the convolution kernel parameters. This indicates the processing results of a hybrid single-rank expert group that shares tasks. This indicates the processing result of a task-specific hybrid single-rank expert group.

6. The multi-task dense prediction method based on hybrid single-rank experts according to claim 4, characterized in that, The channel processing unit consists of a hybrid single-rank expert group combining task sharing and task-specific methods, along with convolutional operations; each expert is composed of... and parameter vector Composition; the processing procedure of the channel processing unit is specifically represented as follows: in, Indicates batch size, This indicates a dimensional change operation. Indicates the first Each token characteristic, This represents the matrix multiplication operation.

7. The multi-task dense prediction method based on hybrid single-rank experts according to claim 1, characterized in that, For the task The prediction module operates as follows: in, Indicates an upsampling operation. For the model to the task The dense prediction results.

8. The multi-task dense prediction method based on hybrid single-rank experts according to claim 1, characterized in that, The multi-task loss joint optimization strategy adopts a weighted multi-task joint loss as the overall optimization objective, and the calculation of the joint loss is expressed as follows: in, Indicates the task index. Indicates the total number of tasks. Indicates the first The weight of the loss term for each task, The overall optimization objective of the model.

9. The multi-task dense prediction method based on hybrid single-rank experts according to claim 8, characterized in that, The loss function is selected based on the task type, including: The semantic segmentation task employs the cross-entropy loss function; The edge detection task employs a balanced cross-entropy loss function. The surface normal estimation task uses the L1 loss function.

10. The multi-task dense prediction method based on hybrid single-rank experts according to claim 1, characterized in that, The performance evaluation adopts evaluation metrics corresponding to the task, including: using the average intersection-union ratio to evaluate the semantic segmentation task, using the root mean square error to evaluate the surface normal estimation task, and using the optimal dataset scale F-value to evaluate the edge detection task.

Images