Auxiliary alignment method for object classification and positioning regression
By optimizing the selection of candidate positive sample points and introducing the location regression branch loss function and the joint branch of classification and positioning, the problems of scarcity of small and medium-sized target positive samples and misalignment of classification and positioning in target detection are solved, and the detection performance and alignment effect of the model are improved.
Patent Information
- Application Number
- CN202410981920.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-08
AI Technical Summary
The existing object detection methods lack positive samples for small and medium-sized targets during training, resulting in poor detection results, and object classification and positioning regression are easily misaligned, resulting in overall performance degradation.
Optimize the candidate positive sample point selection scheme, build the positioning regression branch loss function, and adopt a new dynamic label allocation method and joint branch of classification and positioning to ensure that the model obtains sufficient positive samples in the early stage of training, so as to align object classification and positioning tasks.
By optimizing the selection of positive sample points and introducing joint branches, the model can obtain high scores in both classification and positioning, improving the performance and generalization ability of target detection.
Smart Images

Figure CN118918415A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of target detection, and more specifically, to an auxiliary alignment method for object classification and positioning regression. Background Art
[0002] The currently popular anchor-free object detector maps the object bounding box to the output feature map during training, and selects the feature points within the object bounding box as candidate positive sample points of the object. Although this processing shows good results in many scenarios, during the training process, since large objects occupy more pixel space, sufficient positive samples are usually obtained, while small and medium-sized objects occupy less pixel space and usually suffer from the scarcity of positive samples, or even lack of positive samples sometimes, which leads to poor detection of small and medium-sized objects in practical applications, causing the overall performance of the detector to decline.
[0003] Label assignment methods can be divided into two categories: static and dynamic. In static methods, the training samples of each object remain fixed throughout the training process; while in dynamic methods, the training samples of each object are variable during the training process. Static label assignment methods usually directly select points with a distance from the center of the object bounding box less than a threshold as positive samples from candidate positive sample points, but this method is not applicable to all object shapes. Some studies select candidate positive sample points with predicted IoU greater than the threshold or ranked higher in descending order as positive samples based on the IoU (Intersection over Union) value for evaluating positioning accuracy. This dynamic label assignment method only considers positioning regression, which may result in high positioning accuracy and low classification scores. Although some studies have taken into account both object classification and positioning regression, according to the product value of the exponential power of the positioning accuracy IoU and the classification score, the candidate positive sample points greater than the threshold or ranked higher in descending order are selected as positive samples, but in fact, the product values corresponding to most candidate positive sample points are very small at the beginning of training, and it is easy to miss suitable positive samples at the beginning of training.
[0004] The target detection deep neural network usually decomposes the target detection task into two subtasks: object classification and positioning regression. The pre-trained classification model is usually used as the backbone to extract features. Existing studies have found that the features that are most suitable for object classification are usually inconsistent with the features that are most suitable for positioning regression. There is usually a deviation between the positions of feature points with high classification scores and feature points with high positioning accuracy, which can easily lead to misalignment between object classification and positioning regression, reducing the performance of the target detection deep neural network.
[0005] In summary, the current target detection methods have the following defects: 1. Popular target detectors only select feature points within the object bounding box as candidate positive sample points of the object. The pixel space occupied by small and medium-sized targets is small, and positive samples are easily lacking; 2. Some dynamic label assignment methods only consider positioning regression, which may result in high positioning accuracy and low classification scores; although some methods take into account both object classification and positioning regression, they are prone to miss suitable positive samples in the early stages of training; 3. There is usually a deviation between the positions of feature points with high classification scores and feature points with high positioning accuracy, and object classification and positioning regression are prone to misalignment. Summary of the invention
[0006] In order to overcome the above-mentioned defects in the prior art, the present invention provides an auxiliary alignment method for object classification and positioning regression.
[0007] To achieve the above object, the technical solution of the present invention is as follows:
[0008] Optimize the candidate positive sample point selection scheme;
[0009] At the same time, a positioning regression branch loss function adapted to the optimization candidate positive sample point selection scheme is constructed;
[0010] After completing the selection of candidate positive sample points using the optimized candidate positive sample point selection scheme, a new dynamic label assignment method is used to determine positive samples and negative samples;
[0011] A classification and localization joint branch for predicting the joint score and aligning the classification branch and the localization branch is introduced into the network structure of the model.
[0012] Preferably, the optimization candidate positive sample point selection scheme is specifically as follows:
[0013] If the bounding box of the object intersects with the associated area of the feature map point, the feature map point can be selected as a candidate positive sample point;
[0014] When the candidate positive sample point areas of multiple objects overlap, since a candidate positive sample point can only be assigned to one object, the overlapping point is assigned to the object with the smallest area according to the area of the candidate positive sample point area of the object and is removed from the candidate positive sample points of other objects.
[0015] Preferably, the associated area is defined as:
[0016] The point coordinates (x, y) of the output feature map correspond to the point coordinates of the input image:
[0017]
[0018] where s x andy is the step length, Indicates rounding down;
[0019] Each output feature map point is centered at the corresponding position of the point in the input image and has an area of s x *s y The small area is the associated area.
[0020] Preferably, the positioning regression branch loss function includes positioning loss and regularization term, and the positioning loss includes: distance loss, shape loss and IoU loss, specifically:
[0021] Distance loss:
[0022]
[0023] Shape loss:
[0024]
[0025] IoU Loss:
[0026]
[0027] Among the three loss expressions, B p and B t Represent the predicted bounding box and the true bounding box of the object, respectively, c For B p and B t The center distance, d b For B p and B t The length of the diagonal of the minimum enclosing rectangle, θ x and θ y Respectively represent B p and B t The angle between the center line and the coordinate axis, θ = max(θ x ,θ y ), w p and w t Respectively represent B p and B t Width, h p and h t Respectively represent B p and B t of high;
[0028] The positioning regression branch loss function can be expressed as:
[0029]
[0030] Where N pos is the number of positive samples in the output feature map, is the localization regression loss of the i-th positive sample.
[0031] The expression of the positioning regression loss of the positive sample is:
[0032] L reg =L box +λ range [max(0,l p +r p -1) 2 +max(0,t p +b p -1) 2 ]
[0033] Among them, [max(0,l p +r p -1) 2 +max(0,t p +b p -1) 2 ] is the regularization term, L box is the object positioning loss, L box The expression is:
[0034]
[0035] Among them, the weight
[0036] L reg In the expression of range Used to control the influence of the regularization term, l p ,t p 、r p 、b p is the localization regression branch output of the positive sample point, l p ,t p 、r p 、b p The expressions are:
[0037] l p =x p -x tl
[0038] t p =y p -y tl
[0039] r p =x br -x p
[0040] b p =y br -yp
[0041] Among them, (x p ,y p ) is the coordinate of the positive sample point of the object, (x tl ,y tl ) and (x br ,y br ) are the coordinates of the upper left corner and lower right corner of the object bounding box, and after normalizing the coordinates, l p ,t p 、r p 、b p The value is between -1 and 1. p +r p ,t p +b p The value is between 0 and 1.
[0042] Preferably, the new dynamic allocation method comprises:
[0043] The classification score and positioning accuracy of the predicted candidate positive sample points of the object are jointly evaluated to obtain a joint score, and the joint scores corresponding to all candidate positive sample points are sorted according to size, among which the largest top M k The candidate positive sample points corresponding to the joint scores are taken as positive samples, and the points other than the candidate positive sample points in the output feature map are taken as negative samples. During the training process, the number of positive samples is decreased as the number of training rounds increases.
[0044] Preferably, the classification score and positioning accuracy of the predicted candidate positive sample points of the combined object are combined to obtain a combined score, and the combined scores are sorted by size as follows:
[0045] Define the joint score S joint for:
[0046]
[0047] Where S cls is the classification score of the object, L box is the object positioning loss, 1-L box is the object positioning accuracy, α, β∈[0,1], α+β=1, α and β can adjust the importance of classification and positioning respectively;
[0048] After obtaining the joint score, the joint scores corresponding to all candidate positive sample points are arranged in descending order.
[0049] Preferably, the number of positive samples is decreased as the number of rounds increases during the training process as follows:
[0050] Let the number of positive samples of the object in the kth round be Mk for:
[0051] M k =(M0-1)e -kr +1
[0052] Among them, M0 is the number of candidate positive samples; parameter r>0 can control the decreasing speed of the number of positive samples. The larger r is, the smaller the number of positive samples M is. k The faster it falls.
[0053] Preferably, the loss of the classification and positioning joint branch includes a focal loss Focal Loss and a smooth L1 loss Smooth L1 Loss, and the focal loss Focal Loss and the smooth L1 loss Smooth L1 Loss are specifically:
[0054]
[0055]
[0056] Among them, FL and L1 represent Focal Loss and Smooth L1 Loss respectively;
[0057] α f ∈[0,1], used to suppress the imbalance in the number of positive and negative samples;
[0058] γ ≥ 0, used to control the imbalance in the number of difficult and easy samples;
[0059] u>0, used to control the smoothness of the zero point;
[0060] d=S joint -p, in the training process of the joint classification and localization branch, p is the predicted value of the output feature map point, S moint is the joint score of classification and localization.
[0061] Preferably, the loss function L of the loss of the joint classification and positioning branch is J for:
[0062]
[0063] Among them, λ L1 is the weight, M is the number of points in the output feature map, is the Smooth L1 Loss of the mth point; N pos and N neg are the number of positive samples and negative samples in the output feature map, respectively. is the Focal Loss of the i-th positive sample, is the Focal Loss of the jth negative sample.
[0064] Preferably, after introducing the classification and positioning joint branch, the total loss of the final model is:
[0065] L=λ C L C +λ R L R +λ J L J
[0066] Among them, λ C , R and λ J is a hyperparameter; L C It is the sum of the average classification losses of positive and negative samples, which can be calculated by the focal loss expression.
[0067] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:
[0068] The present invention optimizes the candidate positive sample point selection scheme and constructs a positioning regression branch loss function adapted thereto, adopts a new dynamic label allocation method, and introduces a classification and positioning joint branch to ensure that all targets in the input image have candidate positive sample points available for training. The model can obtain a sufficient number of positive samples for learning in the early stage of training, and can directly obtain target detection results with high scores in both classification and positioning. Without changing the original object classification branch and positioning regression branch structure, the object classification task and the positioning task are aligned. BRIEF DESCRIPTION OF THE DRAWINGS
[0069] Figure 1 Schematic diagram of the candidate positive sample point selection scheme.
[0070] Figure 2 This is a graph showing the relationship between the number of positive samples decreasing with the rounds.
[0071] Figure 3 for l p ,t p 、r p 、b p Definition diagram of .
[0072] Figure 4 For L box Calculation diagram.
[0073] Figure 5 This is a model detection diagram of a dog in the embodiment.
[0074] Figure 6 The figure is a network structure diagram after adopting the method of the present invention. DETAILED DESCRIPTION
[0075] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.
[0076] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.
[0077] Example 1
[0078] This embodiment includes the following steps:
[0079] S1: Please refer to Figure 6 , the network structure of the model is as follows Figure 6 As shown in the figure, it mainly includes the backbone network, feature fusion network and detection head. The backbone network and feature fusion network are designed to extract and fuse multi-scale features of the image. The detection head contains classification branch, localization regression branch and classification and localization joint branch.
[0080] If the output feature map size is H×W and the number of target categories is C, the output dimension of the classification branch is H×W×C, the output dimension of the localization and regression branch is H×W×4, and the output dimension of the joint classification and localization branch is H×W×1.
[0081] In this embodiment, each feature graph point in the classification branch is a C-dimensional vector, corresponding to C target categories, and the output value is limited to between 0 and 1 through the softmax function; each feature graph point in the positioning regression branch is output as a 4-dimensional vector l p ,t p 、r p 、b p , corresponding to the expressions:
[0082] l p =x p -x tl
[0083] t p =y p -y tl
[0084] r p =x br -x p
[0085] b p =y br -y p
[0086] Among them, (xp ,y p ) is the coordinate of the positive sample point of the object, (x tl ,y tl ) and (x br ,y br ) are the coordinates of the upper left corner and lower right corner of the object’s bounding box, respectively.
[0087] The classification and positioning joint branch outputs a joint score S for each feature map point joint , the output value is limited to between 0 and 1 through the sigmoid function.
[0088] S2: During the model training phase, the candidate positive sample point selection scheme in the embodiment of the present invention can ensure that each object has positive samples for learning.
[0089] See also Figure 1 , if the associated area of the output feature map point intersects with the object bounding box, then the point is a candidate positive sample point of the object. The definition of the associated area is:
[0090] The point coordinates (x, y) of the output feature map correspond to the point coordinates of the input image:
[0091]
[0092] where s x and y is the step length, Indicates rounding down; each output feature map point is centered on the corresponding position of the point in the input image and has an area of s x *s y The small area is the associated area.
[0093] When the candidate positive sample points of multiple objects overlap, the overlapping points are assigned to the object with a smaller area according to the area of the candidate positive sample points of the object, and the overlapping points are removed from the candidate positive sample points of other objects. Therefore, even for small objects, it is guaranteed that there is at least one candidate positive sample point.
[0094] S3: During the model training phase, see Figure 6 , define the joint score S of classification and localization joint for:
[0095]
[0096] Among them, S cls is the classification score of the object, L box is the object positioning loss, 1-L box is the positioning accuracy of the object. α and β can adjust the importance of classification and positioning respectively. α, β∈[0,1], α+β=1.
[0097] The classification score and positioning accuracy of the candidate positive sample points are jointly evaluated to obtain the joint score S joint , sorted in descending order according to the joint score, and then the number of positive samples in the current training round is determined according to the following formula, specifically:
[0098] Let the number of positive samples of the object in the kth round be M k for:
[0099] M k =(M0-1)e -kr +1
[0100] Where M0 is the number of candidate positive samples; parameter r>0 can control the decreasing speed of the number of positive samples. Figure 2 , the larger r is, the number of positive samples M k The faster it falls.
[0101] After determining the number of positive samples in the current training round, select the candidate positive sample points that rank higher in the joint score ranking as positive samples, and the points other than the candidate positive sample points of the selected target in the output feature map are regarded as negative samples of the selected target. This dynamic label assignment method can continuously optimize the quality of training samples, ensure that each target has at least one positive sample, and help improve the generalization ability and detection performance of the model.
[0102] S4: During the model training phase, the loss of the classification branch usually uses Focal Loss. The average classification loss of positive and negative samples can be calculated according to the above FocalLoss formula. The loss of the classification branch L C is the sum of the average classification losses of positive and negative samples;
[0103] The positioning regression branch loss function includes positioning loss and regularization term. The positioning regression branch loss function includes distance loss, shape loss and IoU loss, which are as follows:
[0104] Distance loss:
[0105]
[0106] Shape loss:
[0107]
[0108] IoU Loss:
[0109]
[0110] Among the three loss expressions, B p and B t Represent the predicted bounding box and the true bounding box of the object, respectively, cFor B p and B t The center distance, d b For B p and B t The length of the diagonal of the minimum enclosing rectangle, θ x and θ y Respectively represent B p and B t The angle between the center line and the coordinate axis, θ = max(θ x ,θ y ), w p and w t Respectively represent B p and B t Width, h p and h t Respectively represent B p and B t of high;
[0111] The specific loss function of the positioning regression branch is:
[0112]
[0113] Where N pos is the number of positive samples in the output feature map, is the localization regression loss of the i-th positive sample.
[0114] The localization regression loss L for each positive sample reg The expression is:
[0115] L reg =L box +λ range [max(0,l p +r p -1) 2 +max(0,t p +b p -1) 2 ]
[0116] Among them, [max(0,l p +r p -1) 2 +max(0,t p +b p -1) 2 ] is the regularization term, L box is the object positioning loss, L box The expression is:
[0117]
[0118] Among them, the weight
[0119] L reg In the expression of range Used to control the influence of the regularization term, l p ,t p 、r p 、b p It is the positioning regression branch output of the above positive sample point.
[0120] The losses of the joint branch of classification and positioning include focal loss Focal Loss and smooth L1 loss Smooth L1 Loss. The focal loss Focal Loss and smooth L1 loss Smooth L1 Loss are specifically:
[0121]
[0122] Among them, FL and L1 represent Focal Loss and Smooth L1 Loss respectively;
[0123] α f ∈[0,1], used to suppress the imbalance in the number of positive and negative samples;
[0124] γ ≥ 0, used to control the imbalance in the number of difficult and easy samples;
[0125] u>0, used to control the smoothness of the zero point;
[0126] d=S joint -p, in the training process of the joint classification and localization branch, p is the predicted value of the output feature map point, S joint is the joint score of classification and localization.
[0127] The loss function L of the joint classification and localization branch J for:
[0128]
[0129] After constructing a new positioning regression branch loss function and introducing the positioning and regression branches, the total loss of the final model can be defined as:
[0130] L=λ C L C +λ R L R +λ J L J
[0131] Among them, λ C , R and λ J is a hyperparameter to ensure that each loss contributes equally to the total loss. The loss L of the classification branchC Calculated according to the above Focal Loss formula.
[0132] S5: In the inference stage, first determine the location of the local maximum of the output of the joint classification and localization branch, and then find the results of the object classification branch and the localization regression branch at that location respectively.
[0133] See also Figure 5 , Figure 5 The results of dog detection by the model after applying the method of the present invention are shown. It can be seen that there are deviations between the positions of the highest joint score point with the joint branch, the highest classification branch score point without the joint branch, and the highest positioning regression IoU point without the joint branch. The highest classification branch score point usually indicates that the confidence that the area contains the target is the highest, the highest positioning regression IoU point indicates the position where the predicted bounding box overlaps the true bounding box to the greatest extent, and the highest joint score point combines the information of the classification branch and the positioning branch to finally determine a comprehensive optimal position. Figure 5 It can be seen that the highest point of the joint score is located at the dog's head, which is consistent with human perception. When humans recognize dogs in images, they usually notice the dog's head first. This result shows that by introducing the joint branch of classification and positioning, the model can not only obtain better detection results, but also be more in line with human intuitive judgment.
[0134] Compared with the prior art, the embodiments of the present invention optimize the candidate positive sample point selection scheme and construct a positioning regression branch loss function adapted thereto, adopt a new dynamic label allocation method, introduce classification and positioning joint branches and other methods, so as to ensure that all targets in the input image have candidate positive sample points available for training. The model can obtain a sufficient number of positive samples for learning in the early stage of training, and can directly obtain target detection results with high scores in both classification and positioning. Without changing the original object classification branch and positioning regression branch structure, the object classification task and the positioning task are aligned.
[0135] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.
Claims
1. An auxiliary alignment method for object classification and positioning regression, characterized in that: The method comprises: Optimize the candidate positive sample point selection scheme; At the same time, a positioning regression branch loss function adapted to the optimization candidate positive sample point selection scheme is constructed; After completing the selection of candidate positive sample points using the optimized candidate positive sample point selection scheme, a new dynamic label allocation method is used to determine positive samples and negative samples; A classification and localization joint branch for predicting the joint score and aligning the classification branch and the localization branch is introduced into the network structure of the model.
2. The auxiliary alignment method for object classification and positioning regression according to claim 1, characterized in that: The optimization candidate positive sample point selection scheme is specifically as follows: If the bounding box of the object intersects with the associated area of the feature point, the feature point can be selected as a candidate positive sample point; When the candidate positive sample point areas of multiple objects overlap, since a candidate positive sample point can only be assigned to one object, the overlapping point is assigned to the object with the smallest area according to the area of the candidate positive sample point area of the object and is removed from the candidate positive sample points of other objects.
3. The auxiliary alignment method for object classification and positioning regression according to claim 2, characterized in that: The definition of the associated region is: The point coordinates (x, y) of the output feature map correspond to the point coordinates of the input image: where s x and y is the step length, Indicates rounding down; Each output feature map point is centered at the corresponding position of the point in the input image and has an area of s x *s y The small area is the associated area.
4. The auxiliary alignment method for object classification and positioning regression according to claim 1, characterized in that: The positioning regression branch loss function includes positioning loss and regularization term. The positioning loss includes: distance loss, shape loss and IoU loss, specifically: Distance loss: Shape loss: IoU Loss: Among the three loss expressions, B p and B t Represent the predicted bounding box and the true bounding box of the object, respectively, c For B p and B t The center distance, d b For B p and B t The length of the diagonal of the minimum enclosing rectangle, θ x and θ y Respectively represent B p and B t The angle between the center line and the coordinate axis, θ = max(θ x ,θ y ), w p and w t Respectively represent B p and B t Width, h p and h t Respectively represent B p and B t of high; The positioning regression branch loss function can be expressed as: Where N pos is the number of positive samples in the output feature map, is the localization regression loss of the i-th positive sample. The expression of the positioning regression loss of the positive sample is: L reg =L box +λ range [max(0,l p +r p -1) 2 +max(0,t p +b p -1) 2 ] Among them, [max(0,l p +r p -1) 2 +max(0,t p +b p -1) 2 ] is the regularization term, L box is the object positioning loss, L box The expression is: Among them, the weight λ IoU , L reg In the expression of range Used to control the influence of the regularization term, l p ,t p 、r p , b p is the localization regression branch output of the positive sample point, l p ,t p 、r p , b p The expressions are: l p =x p -x tl t p =and p -and tl r p =x br -x p b p =and br -and p Among them, (x p ,y p ) is the coordinate of the positive sample point of the object, (x tl ,y tl ) and (x br ,y br ) are the coordinates of the upper left corner and lower right corner of the object bounding box, and after normalizing the coordinates, l p ,t p 、r p , b p ∈[-1,1], l p +r p ,t p +b p ∈[0,1].
5. The auxiliary alignment method for object classification and positioning regression according to claim 1, characterized in that: The new dynamic allocation method includes: The classification score and positioning accuracy of the predicted candidate positive sample points of the object are jointly evaluated to obtain a joint score, and the joint scores corresponding to all candidate positive sample points are sorted according to size, among which the largest top M k The candidate positive sample points corresponding to the joint scores are taken as positive samples, and the points other than the candidate positive sample points in the output feature map are taken as negative samples. During the training process, the number of positive samples is decreased as the number of training rounds increases.
6. The auxiliary alignment method for object classification and positioning regression according to claim 5, characterized in that: The classification score and positioning accuracy of the joint object candidate positive sample point prediction are combined to obtain a joint score, and the joint score is sorted by size as follows: Define the joint score S joint for: Where S cls is the classification score of the object, L box is the object positioning loss, 1-L box is the object positioning accuracy, α, β∈[0,1], α+β=1, α and β can adjust the importance of classification and positioning respectively; After obtaining the joint score, the joint scores corresponding to all candidate positive sample points are arranged in descending order.
7. The auxiliary alignment method for object classification and positioning regression according to claim 5, characterized in that: The number of positive samples is decreased in accordance with the increasing number of rounds during the training process as follows: Let the number of positive samples of the object in the kth round be M k for: M k =(M0-1)e -kr +1 Among them, M0 is the number of candidate positive samples; parameter r>0 can control the decreasing speed of the number of positive samples. The larger r is, the smaller the number of positive samples M is. k The faster it falls.
8. The auxiliary alignment method for object classification and positioning regression according to claim 1, characterized in that: The loss of the classification and positioning joint branch includes focal loss Focal Loss and smooth L1 loss Smooth L1 Loss, and the focal loss Focal Loss and smooth L1 loss Smooth L1 Loss are specifically: Among them, FL and L1 represent Focal Loss and Smooth L1 Loss respectively; α f ∈[0,1], used to suppress the imbalance in the number of positive and negative samples; γ ≥ 0, used to control the imbalance in the number of difficult and easy samples; u>0, used to control the smoothness of the zero point; d=S joint -p, in the training process of the joint classification and localization branch, p is the predicted value of the output feature map point, S joint is the joint score of classification and localization.
9. The auxiliary alignment method for object classification and positioning regression according to claim 1, characterized in that: The loss function L of the joint classification and localization branch is J for: in, is the weight, M is the number of points in the output feature map, is the Smooth L1 Loss of the mth point; N pos and N neg are the number of positive samples and negative samples in the output feature map, respectively. is the Focal Loss of the i-th positive sample, is the Focal Loss of the jth negative sample.
10. The auxiliary alignment method for object classification and positioning regression according to any one of claims 1 to 9, characterized in that: After the introduction of the joint classification and localization branch, the total loss of the final model is: L=λ C L C +λ R L R +λ J L J Among them, λ C , R and λ J is a hyperparameter; L C It is the sum of the average classification losses of positive and negative samples, which can be calculated by the focal loss FocalLoss expression.
