Rotating target detection method based on task decoupling and adaptive point set strategy

Abstract

The present invention discloses a rotating target detection method based on task decoupling and adaptive point set strategy, the method comprising the following steps: Step 1: designing an anchor-free rotating target detection network based on an adaptive point set strategy, mining the deep geometric information of targets in any direction, and extracting the angle information used for rotating target frame positioning; Step 2: designing a detection feature decoupling network, splitting regression and classification into two sub-networks in terms of feature space and network parameters, suppressing the inconsistent feature sensitivity problem between regression tasks and classification tasks, and realizing the detection and positioning of remote sensing targets in complex scenes. The method realizes remote sensing target detection in various complex scenes by enhancing the model's ability to extract deep spatial geometric information and basic texture features, providing strong support for military and civilian applications such as air traffic control, maritime rescue and port management.

Description

Technical Field

[0001] The invention belongs to the field of target detection and detection technology, and relates to a rotating target detection method in a remote sensing image, and specifically to a rotating target detection method based on detection feature decoupling and an adaptive point set strategy. Background Art

[0002] Object detection in remote sensing images aims to locate and classify remote sensing targets. It is a key technology in many military and civilian applications such as air traffic control, maritime rescue and port management. Remote sensing target detection methods are usually divided into two categories: horizontal box target detection and rotation box target detection. Directly applying the horizontal bounding box detection method to rotation target detection will lead to serious deviations between the generated predicted bounding box and the real target bounding box object. In addition, due to the long-distance imaging characteristics of remote sensing, the target texture details are seriously missing, and the targets are mostly small-scale morphology. The negative impact of the position deviation of the predicted box will be magnified in remote sensing tasks.

[0003] Most of the current rotation object detection methods are anchor-based detectors. Usually, these detection methods densely construct anchor boxes on feature maps, and then regress the offset between the target box and the anchor box to achieve accurate positioning of the object. Compared with the horizontal box object detection task, rotation object detection requires additional learning of the angle of the object, that is, the rotation object annotation method mainly uses the center, width, height and angle to describe the rotated object box. There are many different ideas for the regression of the current mainstream angle. For example, most detectors that rely on the two-stage R-CNN framework use RPN to generate rotated regions of interest (rRoIs), and then use the designed RoI pooling operator to extract features from these RoIs. ROI-Transformer applies spatial transformation on the ROI and learns the transformation parameters under the supervision of the rotated object box (OBB) annotation. ReDet explicitly encodes rotation equivalence and rotation invariance in the first stage of the detector. For these anchor-based detectors, the scanning process of the anchor box is very time-consuming. To solve this problem, CenterNet proposed to detect the center point of the target through a heat map and regress the width and height of the target box at the same time. This method can achieve faster convergence and reasoning speed than horizontal box target detection with a small loss of detection accuracy. Although the above methods can realize the learning of angle information, due to the different measurement units of angle and other parameters, the joint convergence of these parameters may lead to low training efficiency and introduce new errors. In addition, most existing methods focus on improving the ability to extract refined features of the target as the research focus of target detection, and rarely consider the problem of reduced classification ability caused by the inconsistency of prediction of classification and regression tasks in the general detection architecture. In response to the above problems, some scholars proposed to decompose the detection network into two independent parameter branches to learn different features separately. For example, networks such as RefineDet use deformable convolution (DCN) to extract aligned task features in the refined anchor box. However, the feature sensitive areas of classification and regression tasks are obviously different. How to map them to the same candidate position has not been accurately answered, which also greatly limits the performance and efficiency of remote sensing target detection. Therefore, it is urgent to design and optimize the network structure of the remote sensing target detection model in a targeted manner to address the problems in remote sensing target detection tasks, such as the difficulty in learning the rotated target box and the inconsistent sensitivity of sub-task features. Summary of the invention

[0004] In order to solve the difficult problems of rotating target frame learning and inconsistent sensitivity of subtask features in remote sensing target detection, the present invention provides a rotating target detection method based on task decoupling and adaptive point set strategy. The method realizes remote sensing target detection in various complex scenes by enhancing the model's ability to extract deep spatial geometric information and basic texture features, providing strong support for military and civilian applications such as air traffic control, maritime rescue and port management.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] A rotating target detection method based on task decoupling and adaptive point set strategy includes the following steps:

[0007] Step 1: Design an anchor-free rotation target detection network based on an adaptive point set strategy to mine the deep geometric information of targets in any direction and extract the angle information used for rotating target box positioning;

[0008] Step 2: Design a detection feature decoupling network to split regression and classification into two sub-networks in terms of feature space and network parameters, suppress the inconsistent feature sensitivity between regression tasks and classification tasks, and realize the detection and positioning of remote sensing targets in complex scenarios.

[0009] Compared with the prior art, the present invention has the following advantages:

[0010] (1) In order to solve the problems in remote sensing target detection, such as large joint convergence error of rotating targets, time-consuming anchor frame scanning process, inconsistent sensitivity of subtask features, etc., the present invention proposes an end-to-end rotating target detection method based on task decoupling and adaptive point set strategy. By enhancing the model's ability to extract deep spatial geometric information and basic texture features, remote sensing target detection can be achieved in various complex scenarios, providing strong support for military and civilian applications such as air traffic control, maritime rescue and port management.

[0011] (2) Aiming at the problems of difficulty in learning the rotated target frame and time-consuming anchor frame scanning process, the present invention designs an adaptive point set target detection network based on anchor-free architecture, which mainly includes a target representation module based on adaptive point sets and a corner point regression module based on optimal transmission strategy. The center point position of the target is obtained in the form of a heat map, and the rotated target represented in the form of an adaptive point set is realized through a deformable convolutional network. At the same time, the sinkhorn algorithm is used to realize the regression problem between the target point set predicted by the model and the target true value corner point. This method extracts the semantic information and geometric information of the target through point set representation, and expands the rotated target detection capability of the baseline architecture without introducing additional angle regression errors.

[0012] (3) In order to solve the problems of inconsistent sensitivity of sub-task features and difficulty in accurate target classification in remote sensing targets, the present invention designs a remote sensing target detection network based on detection feature decoupling. The task-sensitive spatial regions are divided by a point set target representation method, and a task-sensitive target detection sub-network for feature space and network parameters is constructed to improve the accuracy of remote sensing target detection. This can realize accurate detection of high-precision remote sensing targets in complex ground-air environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The overall flow chart of the rotating target detection method based on task decoupling and adaptive point set strategy;

[0014] Figure 2 Schematic diagram of the detection head structure based on detection feature decoupling and point set target representation;

[0015] Figure 3 Point set distance representation and cost matrix for the sinkhorn transformation function;

[0016] Figure 4 is a Gaussian distribution based on an adaptive point set;

[0017] Figure 5 Typical target detection results for DOTA public dataset images. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0019] The present invention provides a fine-grained target detection method based on detection feature decoupling and adjacent branch feature fusion, such as Figure 1 As shown, the method comprises the following steps:

[0020] Step 1: Considering that the angle of the rotated target box is measured differently from other parameters, directly converging the angle and other parameters may introduce additional errors and reduce the detection performance of remote sensing targets in the image. Therefore, an adaptive point set target detection method based on the anchor-free architecture is proposed to achieve the detection capability of rotated targets while ensuring that no additional regression errors are introduced and the inference speed is not reduced as much as possible. That is: an adaptive point set target detection network is proposed to extract the semantic information and geometric information of the target through point set representation. The adaptive point set target detection network mainly includes a target representation module based on an adaptive point set and a corner regression module based on an optimal transmission strategy. The target representation module aims to use a heat map method and a deformable convolutional network to obtain an adaptive point set that represents the target area. The corner regression module uses an adaptive point set as input and establishes a mapping relationship between the adaptive point set and the target corner point through a conversion function. The specific steps are as follows:

[0021] Step 1-1: For the input remote sensing image, use the ResNet-101 feature extraction network to obtain a feature map with a downsampling level of 32 times.

[0022] Step 1-2: Perform 8 times upsampling through an upsampling module to restore the resolution of the output feature map to 1 / 4 of the original image. To improve the efficiency of network optimization, a residual structure is used to skip-connect the multi-level features and input features.

[0023] Step 1-3: After the upsampling module, two independent prediction modules are connected to predict the target center point position represented by a heat map and the center point deviation during the image scaling process.

[0024] Step 1-4: Take the predicted target center point as input, obtain the coordinate point position offset through the deformable convolutional network, and construct an adaptive point set consisting of n coordinate points (see step 2-2 for specific operation steps).

[0025] Steps 1-5: After obtaining the adaptive point set of the target, in order to facilitate the application of the point set in the detection model training and reasoning stages, three different conversion functions are introduced: sinkhorn conversion function, Gaussian conversion function and MinAreaRect function, among which: the sinkhorn conversion function designs the mapping relationship between the adaptive point set and the target corner point, and the Gaussian conversion function converts the adaptive point set into a rotated target frame. The two work together to converge the parameters in the model training stage. The MinAreaRect function in the OpenCV function library is applied to the reasoning stage of the model to obtain the standard rotation prediction frame through the adaptive point set. The specific steps are as follows:

[0026] Step 1-5-1: For the sinkhorn conversion function, such as Figure 3As shown in the figure, if the adaptive point set is A and the target corner point set is B, the Euclidean distance between the two points in the A and B sets is calculated. The distance set is represented by D, and the elements in the set D are represented by D i,j Indicates. The formula is written as:

[0027] D i,j =Distance(A i ,B j )i=1,2,3,4…,N and j=1,2,3,4

[0028] Where N represents the number of predicted points in the adaptive point set A, Distance represents the Euclidean distance calculation function, i refers to the predicted point in the adaptive point set A, and j refers to the target corner point in the target corner point set B.

[0029] Step 1-5-2: Construct the distance set D into a two-dimensional matrix D N×4 , and used as the initialization parameter of the cost matrix in the sinkhorn transformation function, then the entire matrix is ​​indexed, and the rows and columns of the matrix are normalized respectively; during the training process of the model, the normalization operation of the transformation matrix is ​​repeated continuously, and finally the transformation matrix of the adaptive point set and the target corner point is obtained.

[0030] Step 1-5-3: Figure 4 As shown, the coordinates of the predicted points in the adaptive point set A are set to the two-dimensional variable a i , calculate the probability density function of the adaptive point set A under Gaussian distribution, and use the probability density function as the position optimization target of the Gaussian transformation function.

[0031] Step 1-5-4: The above-mentioned transformation matrix and probability density function are jointly optimized during the training process.

[0032] Step 1-5-5: During the inference process, use the MinAreaRect function to convert the predicted adaptive point set into a standard rotated rectangular box.

[0033] Step 2: Considering the problem of inconsistent task feature sensitivity in target detection, that is, the classification task and regression task focus on different feature areas during the detection process, feature distribution coupling and network parameter sharing will lead to mutual constraints between classification and regression tasks, reducing the model's model detection and target positioning capabilities. Therefore, a network structure with task-level spatial feature decoupling is proposed. Combined with the point set target representation method, a task-sensitive target detection sub-network for feature space and network parameters is constructed while ensuring task consistency and relevance as much as possible, to improve the accuracy of fine-grained aircraft detection and achieve high-precision aircraft model detection in complex ground-to-air environments. Among them, the task feature decoupling network mainly designs two main modules: the classification feature sampling module (Sampling Classification Features Module, SCF) and the regression feature sampling module (SamplingRegression Features Module, SRF). Figure 2 As shown, the specific steps are as follows:

[0034] Step 2-1: Obtain the fused feature map generated by the neighboring feature fusion network, and inject it into the classification feature sampling module (SCF module) and the regression feature sampling module (SRF module) respectively.

[0035] Step 2-2: Use the classification feature sampling module (SCF module) to extract feature areas that are sensitive to the classification task and enhance the ability of the classification network to extract various discriminative features. This module innovatively designs a subnetwork for the classification task, and uses point sets to represent important semantic positions for the target category, models the sampling network of classification features, and enables the module to have adaptive feature semantic extraction capabilities for target classification tasks. The specific structure mainly includes 3 3×3 convolutional layers, 2 1×1 convolutional layers, and 1 3×3 deformable convolutional layer. The specific discrimination steps are:

[0036] Step 2-2-1: For the input fusion feature map, three layers of convolution, batch normalization, and ReLU activation function are used to operate the feature map through three 3×3 convolution layers to achieve classification feature analysis. The formula is written as:

[0037] F cls =ReLU(BN(conv 3×3 (F)))

[0038] In the formula, conv 3×3 , BN and ReLU represent 3×3 convolution, batch normalization, and ReLU activation function operations respectively; F cls It represents the feature map obtained after convolution processing, and F represents the input feature map of the SCF module.

[0039] Step 2-2-2: Compress the high-dimensional feature map into 2×n dimensions through a 1×1 convolution layer, where n represents the number of biased coordinate points in the point set. Each dimension of the convolution stores the position bias of the biased coordinate point. The formula is written as:

[0040] offset cls =conv 1×1 (F cls )

[0041] In the formula, offset cls Represents the position bias in the classification task, conv 3×3 Represents a 3×3 convolution operation.

[0042] Step 2-2-3: Set the offset cls As the bias input of the 3×3 deformable convolutional layer, F cls As the feature map input, the classification feature extraction after bias is realized, and the formula is written as:

[0043] F output =dconv(F cls ,offset cls )

[0044] Where dconv represents the deformable convolution operation, F output Represents a classification feature map.

[0045] Step 2-2-4: Compress the high-dimensional feature map into m dimensions through a 1×1 convolutional layer, where m represents the total number of categories for the final classification. The formula is written as:

[0046] CLS=conv 1×1 (F output )

[0047] In the formula, CLS represents the final output of the classification branch, conv 1×1 Represents a 1×1 convolution operation.

[0048] Step 2-3: Use the regression feature sampling module (SRF module) to extract feature areas that are sensitive to the regression task and enhance the ability of the regression network to extract border positioning features. This module innovatively designs a subnetwork for regression tasks, uses point sets to extract the boundary position of the target area, and models the sampling network of regression features, so that the module has the ability to adaptively locate the boundary of the target area. The specific structure mainly includes 3 3×3 convolutional layers, 2 1×1 convolutional layers, and 1 3×3 deformable convolutional layer. The specific discrimination steps are:

[0049] Step 2-3-1: For the input fusion feature map, three layers of convolution, batch normalization, and ReLU activation function are used to operate the feature map through three 3×3 convolution layers to achieve regression feature parsing. The formula is written as:

[0050] F init =ReLU(BN(conv 3×3 (F)))

[0051] In the formula, conv, BN and ReLU represent convolution, batch normalization and ReLU activation function operations respectively; F init Represents the feature map obtained after convolution processing.

[0052] Step 2-3-2: Compress the high-dimensional feature map into 2×n dimensions through a 1×1 convolution layer, where n represents the number of offset coordinate points in the point set. Each dimension of the convolution stores the position offset of the offset coordinate point. The formula is written as:

[0053] offset init =conv 1×1 (F init )

[0054] In the formula, offset init Represents the position bias in the classification task.

[0055] Step 2-3-3: Set the offset init Add the n initial coordinates in the point set, and map the step size of the fusion feature map of this layer to get the corresponding coordinates pts on the original image. int :

[0056] pts int =convert(offset init )

[0057] In the formula, convert represents the coordinate mapping between the fused feature map and the original image.

[0058] Step 2-3-4: Set the offset init As the bias input of the 3×3 deformable convolutional layer, F init As the feature map input, the regression feature extraction after bias is realized, and the formula is written as:

[0059] F output =dconv(F init ,offset init )

[0060] Where dconv represents the deformable convolution operation, F refine Represents a classification feature map.

[0061] Step 2-3-5: Compress the high-dimensional feature map into 2×n dimensions through a 1×1 convolution layer, where n represents the number of biased coordinate points in the point set. Each dimension of the convolution stores the position bias of the biased coordinate point. The formula is written as:

[0062] offset refine =conv 1×1 (F output )

[0063] In the formula, offset refine Represents the position bias in the classification task.

[0064] Step 2-3-6: pts int with offset refine Add them together and get the corresponding coordinates on the original image according to the step size mapping of the fused feature map of this layer:

[0065] pts refine =convert(pts int ,offset refine )

[0066] In the formula, convert represents the coordinate mapping between the fused feature map and the original image, pts refine Represents the final output of the regression branch.

[0067] Figure 5 The detection results of typical ships and aircraft in the DOTA public dataset are given. From the detection results, it can be seen that in complex port and airport scenarios, the detection method proposed in the present invention can correctly detect small and medium-sized ships and aircraft targets in any direction without generating false alarms.

Claims

1. A rotating target detection method based on task decoupling and adaptive point set strategy, characterized in that The method comprises the following steps: Step 1: Design an anchor-free rotation target detection network based on an adaptive point set strategy to mine the deep geometric information of targets in any direction and extract the angle information used for rotating target box positioning; Step 2: Design a detection feature decoupling network, split regression and classification into two sub-networks in terms of feature space and network parameters, suppress the inconsistent feature sensitivity between regression tasks and classification tasks, and realize the detection and positioning of remote sensing targets in complex scenes. The specific steps are as follows: Step 2-1: Obtain the fused feature map generated by the neighboring feature fusion network, and inject it into the classification feature sampling module and the regression feature sampling module respectively; Step 2-2: Use the classification feature sampling module to extract feature areas that are sensitive to the classification task and enhance the ability of the classification network to extract various discriminant features. The specific steps are as follows: Step 2-2-1: For the input fusion feature map, three layers of convolution, batch normalization, and ReLU activation function are used to operate the feature map through three 3×3 convolution layers to achieve classification feature analysis. The formula is written as: F cls =ReLU(BN(conv 3×3 (F))) In the formula, conv 3×3 , BN and ReLU represent 3×3 convolution, batch normalization, and ReLU activation function operations respectively; F cls represents the feature map obtained after convolution processing, and F represents the input feature map of the SCF module; Step 2-2-2: Compress the high-dimensional feature map into 2×n dimensions through a 1×1 convolution layer, where n represents the number of biased coordinate points in the point set. Each dimension of the convolution stores the position bias of the biased coordinate point. The formula is written as: offset cls =conv 1×1 (F cls ) Where, offset cls Represents the position bias in the classification task, conv 3×3 Represents a 3×3 convolution operation; Step 2-2-3: Set the offset cls As the bias input of the 3×3 deformable convolutional layer, F cls As the feature map input, the classification feature extraction after bias is realized, and the formula is written as: F output =dconv(F cls ,offset cls ) Where dconv represents the deformable convolution operation, F output Represents a classification feature map; Step 2-2-4: Compress the high-dimensional feature map into m dimensions through a 1×1 convolutional layer, where m represents the total number of categories for the final classification. The formula is written as: CLS=conv 1×1 (F output ) In the formula, CLS represents the final output of the classification branch, conv 1×1 Represents a 1×1 convolution operation; Step 2-3: Use the regression feature sampling module to extract feature areas that are sensitive to the regression task and enhance the ability of the regression network to extract border positioning features. The specific steps are as follows: Step 2-3-1: For the input fusion feature map, three layers of convolution, batch normalization, and ReLU activation function are used to operate the feature map through three 3×3 convolution layers to achieve regression feature parsing. The formula is written as: F init =ReLU(BN(conv 3×3 (F))) In the formula, conv, BN and ReLU represent convolution, batch normalization and ReLU activation function operations respectively; F init Represents the feature map obtained after convolution processing; Step 2-3-2: Compress the high-dimensional feature map into 2×n dimensions through a 1×1 convolution layer, where n represents the number of biased coordinate points in the point set. Each dimension of the convolution stores the position bias of the biased coordinate point. The formula is written as: offset init =conv 1×1 (F init ) Where, offset init Represents the position bias in the classification task; Step 2-3-3: Set the offset init Add the n initial coordinates in the point set, and map the step size of the fusion feature map of this layer to get the corresponding coordinates pts on the original image. int : pts int =convert(offset init ) In the formula, convert represents the coordinate mapping between the fused feature map and the original image; Step 2-3-4: Set the offset init As the bias input of the 3×3 deformable convolutional layer, F init As the feature map input, the regression feature extraction after bias is realized, and the formula is written as: F output =dconv(F init ,offset init ) Where dconv represents the deformable convolution operation, F refine Represents a classification feature map; Step 2-3-5: Compress the high-dimensional feature map into 2×n dimensions through a 1×1 convolution layer, where n represents the number of offset coordinate points in the point set. Each dimension of the convolution stores the position offset of the offset coordinate point. The formula is written as: offset refine =conv 1×1 (F output ) Where, offset refine Represents the position bias in the classification task; Step 2-3-6: pts int with offset refine Add them together and get the corresponding coordinates on the original image according to the step size mapping of the fused feature map of this layer: pts refine =convert(pts int ,offset refine ) In the formula, convert represents the coordinate mapping between the fused feature map and the original image, pts refine Represents the final output of the regression branch.

2. The rotating target detection method based on task decoupling and adaptive point set strategy according to claim 1 is characterized in that The anchor-free rotation target detection network based on the adaptive point set strategy includes a target characterization module based on the adaptive point set and a corner point regression module based on the optimal transmission strategy, wherein the target characterization module uses a heat map method and a deformable convolutional network to obtain an adaptive point set that characterizes the target area, and the corner point regression module uses the adaptive point set as input and establishes a mapping relationship between the adaptive point set and the target corner point through a conversion function.

3. The rotating target detection method based on task decoupling and adaptive point set strategy according to claim 1 is characterized in that The specific steps of step 1 are as follows: Step 1-1: For the input remote sensing image, use the ResNet-101 feature extraction network to obtain a feature map with a downsampling level of 32 times; Step 1-2: Perform 8 times upsampling through an upsampling module to restore the resolution of the output feature map to 1 / 4 of the original image. To improve the efficiency of network optimization, a residual structure is used to skip-connect the multi-level features and input features. Step 1-3: After the upsampling module, two independent prediction modules are connected to predict the target center point position represented by a heat map and the center point deviation during the image scaling process; Step 1-4: Take the predicted target center point as input, obtain the coordinate point position offset through the deformable convolutional network, and construct an adaptive point set consisting of n coordinate points; Step 1-5: After obtaining the adaptive point set of the target, in order to facilitate the application of the point set in the detection model training and reasoning stages, three different conversion functions are introduced: sinkhorn conversion function, Gaussian conversion function and MinAreaRect function. Among them: the sinkhorn conversion function designs the mapping relationship between the adaptive point set and the target corner point, and the Gaussian conversion function converts the adaptive point set into a rotated target frame. The two work together to converge the parameters in the model training stage. The MinAreaRect function is applied to the reasoning stage of the model to obtain the standard rotation prediction frame through the adaptive point set.

4. The rotating target detection method based on task decoupling and adaptive point set strategy according to claim 3 is characterized in that The specific steps of steps 1-5 are as follows: Step 1-5-1: For the sinkhorn conversion function, if the adaptive point set is A and the target corner point set is B, calculate the Euclidean distance between the two points in the A and B sets. The distance set is represented by D, and the elements in the set D are represented by D i,j Expressed, the formula is written as: D i,j = Distance(A i , B j ) where i = 1, 2, 3, 4…, N and j = 1, 2, 3, 4 Where N represents the number of predicted points in the adaptive point set A, Distance represents the Euclidean distance calculation function, i refers to the predicted point in the adaptive point set A, and j refers to the target corner point in the target corner point set B; Step 1-5-2: Construct the distance set D into a two-dimensional matrix D N×4 , and used as the initialization parameter of the cost matrix in the sinkhorn transformation function, then the entire matrix is ​​indexed, and the rows and columns of the matrix are normalized respectively; during the training process of the model, the normalization operation of the transformation matrix is ​​repeated continuously, and finally the transformation matrix of the adaptive point set and the target corner point is obtained; Step 1-5-3: Set the coordinates of the prediction points in the adaptive point set A to the two-dimensional variable a i , calculate the probability density function of the adaptive point set A under Gaussian distribution, and use the probability density function as the position optimization target of the Gaussian transformation function; Step 1-5-4: Jointly optimize the above transformation matrix and probability density function during the training process; Step 1-5-5: During the inference process, use the MinAreaRect function to convert the predicted adaptive point set into a standard rotated rectangular box.

5. The rotating target detection method based on task decoupling and adaptive point set strategy according to claim 1 is characterized in that The detection feature decoupling network includes a classification feature sampling module and a regression feature sampling module, wherein: the classification feature sampling module includes 3 3×3 convolutional layers, 2 1×1 convolutional layers, and 1 3×3 deformable convolutional layer; the regression feature sampling module includes 3 3×3 convolutional layers, 2 1×1 convolutional layers, and 1 3×3 deformable convolutional layer.

Images