• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Learning power Gaussian modeling loss for dense rotated object detection in remote sensing images

    2023-11-10 02:16:20YngLIHiningWANGYuqingFANGShengjinWANGZhiLIBitoJIANG
    CHINESE JOURNAL OF AERONAUTICS 2023年10期

    Yng LI, Hining WANG, Yuqing FANG, Shengjin WANG, Zhi LI,Bito JIANG

    a Department of Space Information, Space Engineering University, Beijing 101416, China

    b Beijing Institute of Remote Sensing Information, Beijing 100192, China

    c Department of Electronic Engineering, Tsinghua University, Beijing 100084, China

    KEYWORDS

    Abstract Object detection in Remote Sensing (RS) has achieved tremendous advances in recent years, but it remains challenging for rotated object detection due to cluttered backgrounds, dense object arrangements and the wide range of size variations among objects.To tackle this problem,Dense Context Feature Pyramid Network(DCFPN)and a power α-Gaussian loss are designed for rotated object detection in this paper.The proposed DCFPN can extract multi-scale information densely and accurately by leveraging a dense multi-path dilation layer to cover all sizes of objects in remote sensing scenarios.For more accurate detection while avoiding bottlenecks such as boundary discontinuity in rotated bounding box regression, α-Gaussian loss, a unified power generalization of existing Gaussian modeling losses is proposed.Furthermore, the properties of α-Gaussian loss are analyzed comprehensively for a wider range of applications.Experimental results on four datasets (UCAS-AOD, HRSC2016, DIOR-R, and DOTA) show the effectiveness of the proposed method using different detectors,and are superior to the existing methods in both feature extraction and bounding box regression.

    1.Introduction

    Object detection has developed rapidly in RS and has been applied to marine rescue, disaster prediction, urban monitoring and other areas.Classical detection models use horizontal bounding boxes to locate objects, which includes one-stage object detectors1–3and two-stage detectors.4–6Unlike natural imagery, RS images are recorded by satellites from an aerial perspective, so that objects are densely arranged and appear in arbitrary orientations.Moreover, complex background and large aspect ratio also make horizontal detectors unable to locate the target accurately.By contrast, using oriented bounding boxes can reduce the overlap with background areas and draw the edges of objects precisely.Therefore, it is necessary to conduct researches on rotated object detection for RS images.Current rotated object detection algorithms are designed for different issues.For example, two-stage rotated object detectors7–9are utilized for feature extraction.R3Det10applies one-stage detectors to combine speed and accuracy.Some works11–13are designed for optimizing the position regression of the rotating boxes.Other works14,15are anchor-free frameworks without parameters of anchor boxes.Summarizing the above, current rotating object detection faces two major challenges:

    (1) Effective multi-scale feature extraction.Objects in RS images cover a wide range of size variations due to the spatial resolutions of satellite sensors, also due to the size variation between classes (e.g., airport vs.vehicle)and in the class (e.g., carrier vs.fishing vessel).Therefore, the general structure of feature extraction is difficult to extract features of various RS targets, thus increasing the difficulty of object detection effectively and completely.

    (2) Accurate bounding boxes localization of arbitraryoriented objects.Massive small objects are densely arranged with cluttered backgrounds in remote sensing images, making it difficult to detect specific locations.For large aspect ratios objects in some categories, small angular changes may lead to large Intersection over Union(IoU) deviation during bounding box regression,which requires other metrics to redefine the regression loss.In addition, the definitions of bounding boxes and the periodicity of the angle will also affect the performance of localization.Therefore, there is still a need to further increase the accuracy of positioning for rotating objects during detection.

    In this paper, we provide powerful solutions to the above problems.First,a new feature extraction structure is proposed,which uses a dense multi-path dilation layer to cover all sizes of objects densely in RS images.This structure extracts multi-scale semantic information accurately and efficiently,and performs feature transfer and fusion in deep neural networks.In addition, for more accurate rotated bounding box regression, a new transformation of loss is designed based on existing Gaussian modeling regression losses.Solutions above can be easily integrated into object detector to organize an end-to-end detection network.This also helps improve the detection performance without enhancing the training difficulty.

    To summarize, the key contributions are summarized as follows:

    (1) To enrich the spatial representation of semantic features,we explore a new sub-network called DCFPN.It can expand the scale ranges of receptive fields, and extract feature maps densely and effectively; and therefore, further improves the ability to detect objects of various sizes in RS images.

    (2) We propose α-Gaussian loss,a unified power generalization of existing Gaussian modeling losses including GWD,16KLD,17and KFIoU,18to obtain an accurate and reliable rotated bounding box regression.Experiments show that our approach outperforms existing regression losses based on gaussian distribution, and improves the detection efficiency.

    (3) We analyze the properties of α-Gaussian loss comprehensively.Specially, through theoretical analysis, it is shown that the proposed approach can also be applied to other Gaussian-based regression loss using related statistical distances (such as Bhattacharyya Distance19(BD)).

    (4) We perform extensive experiments on four publicly datasets, UCAS-AOD,20HRSC2016,21DIOR-R22and DOTA.23Experimental results show excellent performance of our approach.

    2.Related work

    2.1.Rotated object detection

    To avoid effects of dense horizontal overlapping bounding boxes and represent target locations accurately,rotated object detection uses rotated bounding boxes to represent arbitraryoriented objects, as shown in Fig.1.ROI Transformer,7ICN,8CAD-Net,24AOOD25and SCRDet9are two-stage rotation detectors focusing on feature extraction and fusion.More recently, oriented R-CNN26designed a light-weight oriented RPN and rotated RoI alignment to promise accuracy and efficiency.DODet27proposed a new representation scheme of oriented objects and a localization-guided head to alleviate the spatial and feature misalignment.R3Det10designed a feature refinement module in single-stage rotation detector to obtain fast speed and better accuracy.Gliding Vertex13and RSDet28accurately described the rotated object by regressing the four vertices.Axis Learning15proposed an anchor-free method to reduce computational complexity in one stage.O2-DNet14located each rotated object by predicting a pair of middle lines of the object.Hou et al.29proposed feature pyramid and feature alignment module based on one-stage cascade detector.RADet,30FFA,31DRN32generated fine feature maps for rotating object, and improve the feature representation ability of network for rotating object.LO-Det33designed new channel aggregation and constraint head structures for lightweight detectors.Although these rotating object detectors have made experimental results on public datasets, some issues still remain.

    The key point of rotated object detection lies in the accurate positioning of large aspect ratios objects in complex environments.To predict the rotated bounding box more accurately,the regression loss needs to be further optimized.

    2.2.Bounding box regression loss

    Most detectors extend l1,l2-norm or their variants as regression loss of bounding box.For instance, Fast R-CNN1uses smooth L1 loss34which independently regresses four variables of the horizontal box, ignoring the correlation between the variables and affecting the accuracy of locating.UnitBox35firstly introduced IoU loss function to regress the prediction box as a whole unit in bounding box prediction.Tychsen-Smith and Petersson36proposed Bounded IoU Loss for object detection to maximize the IoU within the ground truth and the Region of Interest(RoI).A series of IoU-variants loss such as GIoU,37DIoU and CIoU38are subsequently developed for horizontal box regression.Focal-EIOU39and RIoU40pay more attention to high IoU objects.

    These functions cannot be used directly in rotating object detection.Not only because rotation detectors have an additional angle parameter which may lead to boundary discontinuity during angle rotation, but the rotating IoU is indifferentiable as well.From the perspective of regression loss optimization, SCRDet9adds a constant factor calculated by IoU to smooth L1 loss, which is more straight forward and effective to solve the problem of the boundary discontinuity.CSL11changes the regression problem of angular prediction into classification.DCL41proposed a new encoding mechanism that further solved the boundary discontinuity problem introduced by long edges definition of bounding box.CenterMap42converted the oriented bounding box regression into the center-probability-map-prediction issue to recognize the target locations and the background pixels.PIoU Loss is proposed by Chen et al.12to accumulate the contribution of all pixels in internal overlapping and calculate directly,which utilize both angle and IoU in regression,and is applicable to both horizontal and rotated bounding boxes.Zheng et al.43designed a projection method to measure the intersection area and solve the uncertainty of convex aroused by arbitrary rotation.BBAVectors44first regresses boundary-aware vectors and obtains the bounding boxes by detecting the center keypoints of the targets.In addition, to tackle the insensitivity to square-like problem and boundary discontinuity issue in rotating detection fundamentally, a series of regression loss function are generated based on Gaussian distribution.Yang et al.16proposed a differentiable loss based on the Gaussian Wasserstein Distance (GWD) by converting the box in arbitrary direction to a two-dimensional Gaussian distribution.Normalized Wasserstein Distance (NWD) is designed by Xu et al.45to estimate the similarity between bounding boxes by their corresponding Gaussian distributions for tiny object detection.Based on Gaussian distributions, Llerena et al.46also presented a similarity measure using the Hellinger Distance.KLD17calculates the Kullback-Leibler Divergence(KLD) between the Gaussian distributions and converts it to the regression loss form.Compared to previous works, KLD dynamically adjusts the gradients of each parameter in bounding box on the basis of the properties of the object.Yang et al.proposed KFIoU18to utilize Gaussian distribution modeling and mimic the mechanism of SkewIoU inherently by Kalman filter, with no need for additional hyperparameter tuning.

    2.3.Gaussian based metrics for rotated bounding box

    Using different statistical distances to measure Gaussian distribution,we firstly model a rotated bounding box B(x,y,w,h,θ)as a 2-D Gaussian distribution N(m,Σ) by the following formula:

    where R denotes the rotation matrix, and Λ denotes the diagonal matrix of eigen value.Specifically, the center point m is represented by the center coordinates of the box, and the covariance matrix Σ is represented by its width, height and rotation angle in the Gaussian distribution.Thus, the regression of rotated boxes can be changed into the distance measure of two Gaussian distributions, as shown in Fig.2.In the new parameter space, N(mt,Σt) and N(mp,Σp) express Gaussian distribution of ground-truth box and predicted box severally.Different distance metrics are shown as follows.Gaussian Wasserstein Distance16is expressed as:

    Kullback-Leibler Divergence (KLD17):

    KFIoU18uses Kalman filter to obtain the distribution of the intersection area:

    Then the area of corresponding rotating boxes is calculated:

    3.Proposed methods

    3.1.DC-FPN

    Fig.2 Visualization of representing a rotated bounding box B(x,y,w,h,θ) by a two-dimensional Gaussian distribution modeling N(m,Σ).

    Objects in RS images usually have large differences in size, so the feature maps output by neural networks must be able to cover different receptive field scales to extract complete object features.FPN47uses the pyramidal structure of convolutional networks to extract high-level semantics throughout.First, it utilizes the bottom-up structure to extract feature maps at different scales, following the top-down structure to up-sample these feature maps.The up-sampling results are then merged with the corresponding bottom-up layers through lateral connections,and finally it outputs feature maps of higher pyramid levels and thus fuses stronger semantic information.However,for the case of larger input image in remote sensing scenes,the receptive field size of the top feature extraction layer should match the input image size to obtain a desired performance,which requires to stack more layers to ensure that the receptive field sizes are appropriate.Nonetheless, with the deepening of the convolutional layer, there will be a contradiction between the feature maps resolution and the receptive field size, that is, high-level feature maps are always with quite low resolution.At the same time,with the stacking of convolutional layers, there is only elements adding and merging between different receptive fields, but no communication and propagation between semantic information.All of these result in the limitation of detection performance and bring great challenges to the design of feature extraction structure.

    To solve effectively, we design Dense Context Feature Pyramid Network (DCFPN) on the basis of FPN, as shown in Fig.3.The Dense Context (DC) module utilizes multipath dilation layers with different dilation rates to extract rich semantic information in large receptive fields of various sizes.In addition, we introduce dense connections to address the vanishing gradient problem in deep networks, while further merging and propagating multi-scale information to enhance semantic information exchanges.

    The basic FPN outputs the last layer of each stage as C2 to C5 in the bottom-up pyramid.Then connected to the topdown pyramid, the output low-resolution feature maps are sequentially up-sampled and element-wise added with the matching Ci to obtain the corresponding feature map set from P2 to P5.

    For the purpose of extracting underlying semantic information and increasing the receptive field scale, we join the extra{P6, P7} layers into the top-down pyramid based on the official settings of FPN, so that the network can detect more targets of different sizes.

    Moreover, adding a DC module after C5 can obtain larger receptive fields without significantly increasing the computational overhead, which supports the feature transfer and enhances the exchanges of contextual information among the deep network.This structure provides global information for large objects in high resolution images, and is conducive to the feature extraction of dense small objects.Inspired by the dense structure,48the DC module consists of cascade dilated convolutional layers,hence the neurons of intermediate feature maps are capable of encoding semantic information from multiple scales.Besides, neurons in each layer are densely connected to tackle the gradient disappearance issue in deep convolutional networks, while covering the semantic information of multi-scale large receptive fields.Each layer can be described as:

    where Ak,r(?)denotes the convolutions with the filter size k=3 and dilation rates r of 3, 6, 12, 18, 24, which respectively correspond to the densely stacked layer y1to y5.An extra upsampling skip connection U(?) is added in parallel with the dilated convolutional layer to directly transfer more precise locations of features which can be described as:

    where C denotes the feature map extracted by the bottom-up pyramid.Finally, the outputs of the DC module and the C5 layer are attached with 3 × 3 convolutions and layer fusion,then generate P(?) to perform feature fusion.Later, the classification and regression subnetworks cooperate to implement object detection.The above can be expressed as follows:

    Fig.3 Network artitecture for rotated object detection, containing the proposed DCFPN and α-Gaussian losses for rotated bounding box localization.

    where Det(?) represents the object detection process for input image I and object OP,S(?)represents the classification subtask and the regression subtask in our proposed model.

    As shown in Fig.4, the basic FPN model cannot extract enough features for the object of interest.In contrast,DCFPN has achieved improvements in feature extraction, which is more advantageous for acquiring various objects in RS images.The extracted features are more significant, the shape is clearer, and the location is more accurate, which prove that our method obtains more powerful feature information and achieves performance increase.Furthermore,a visual comparison of the feature maps is performed,as shown in Fig.5.FPN may have feature omission in the initial extraction process(see the airplane in Fig.5(b)for details).While the DCFPN module proposed in this paper can highly distinguish the object of interest,suppress the background and noise,and refine the feature extraction effectively.

    3.2.α-Gaussian losses for rotated bounding box regression

    Many kinds of objects appear with large aspect ratios such as bridge, ship, airport and harbor in remote sensing scenarios.Especially after introducing the angle parameter in rotated object detection,small changes in angle,center point position,and aspect ratio will all affect the regression and convergence speed of the network.Thus, the difficulty of object detection precisely is increasing.

    Recent advances in arbitrary object detection convert the ln-norm metric or IoU-based loss within the ground-truth box Bp(xp,yp,wp,hp,θp) and the predicted box Bt(xt,yt,wt,ht,θt)into the calculation between two probability measures N(mp,Σp) and N(mt,Σt).The introduction of Gaussian distribution and probability measures solve the problem of inconsistency between metric and loss, boundary discontinuity and square-like problem.These different Gaussian modeling probability measures can be uniformly expressed as

    There are two ways to transform the probability measures into regression losses.For distance metrics such as GWD16and KLD,17current methods generally use nonlinear functions f(?) to Ddis, and the regression loss can be denoted as:

    where hyperparameter τ modulates the entire loss.For IoU metrics such as KFIoU,18the regression loss is described as:

    Fig.4 Visualizations of pyramid features in FPN and our proposed DCFPN for two kinds of objects(airplane and vehicle,left half and right half).For each kind,first row represents the ground truth and obtained features of P3,P4 and P5 pyramid layer in FPN.The second row represents the final results and features of P3, P4 and P5 pyramid layer obtained by DCFPN.

    Fig.5 Visualization of feature maps.

    where k(?) is the loss concerning function.

    To enhance the performance of object localization, we utilize the Box-Cox transformation49to the loss function based on Gaussian distribution metric.(The Box-Cox transformation has been applied for converting the Cross Entropy (CE)loss into the Generalized Cross Entropy (GCE) loss50to improve the robustness of noisy labels in classification tasks.He et al.51applied it to transform the IoU loss to an α-IoU loss, thus improved the performance of horizontal object detection on multiple datasets benchmark).We perform Box-Cox transformation on the existing Gaussian modeling regression loss function, and simplify the parameters in form to make the regression smoother and more expressive,also easier to train and converge.The power transformation of Gaussian modeling regression loss, α-Gaussian loss is calculated as follows:

    where α>0, and ε represents the regularization term.

    We use Fig.6 to compare IoU loss, Gaussian based losses including GWD, KLD and KFIoU, and the corresponding α-Gaussian losses.Fig.6(a)depicts the change of losses with two different aspect ratio cases.Fig.6(b)demonstrates the effect of the center point offset on different loss function.Fig.6(c)explores the relationship between angle difference and losses.It can be seen from the figures that α-Gaussian losses maintain the monotonicity of losses based on Gaussian distribution.Furthermore,with the changes of aspect ratio,offset and angle difference, the original IoU and Gaussian based losses represented by solid lines change more drastically and are more sensitive to large errors, especially when the predicted value is close to the true value.In most cases, our α-Gaussian loss in dash lines has a more stable and moderate trend compared with the corresponding loss,which makes the network training easier.

    3.3.Properties of α-Gaussian losses

    As α-Gaussian losses may extend to different forms and variants corresponding to Gaussian modeling loss functions.We analyze and summarize the properties theoretically before using.

    (1) Monotonicity.For both distance and IoU metrics loss functions, we let Piand Pjbe the predicted bounding boxes of models Tiand Tj, GT is the ground truth.The following deduction can be drawn:

    By calculating min LGuassianand min Lα-Guassian, the corresponding optimal solution of the model can be obtained.The above inference indicates that Lα-Guassianis consistent with LGuassianin monotonicity.

    (2) Scale invariance.For LGuassianwith scale invariance such as KLD, the transformed Lα-Guassianstill has the affine invariance, which can be described as follows:

    Fig.6 Comparison of different losses in three cases of rotating boxes.

    Hence, the scale invariance of KLD is proved.

    (3) Horizontal detection task.For the horizontal detection task, we set the rotation angle term in parameters to 0 after using the Gaussian modeling and calculating the bounding box.Then,the loss function based on distance metric is approximately comparable to the l2-norm loss.Meanwhile, the IoU-based loss function in Gaussian distribution has not changed and can also be used directly.As a result,the α-Gaussian losses are also applicable to the regression loss in horizontal detection task.

    (4) Universality.To demonstrate the universality of our power transformation of Gaussian modeling regression loss, we introduced a new Bhattacharyya distance(BD19).In our definition of Gaussian distributions, the Bhattacharyya distance is defined as:

    Then we apply the power transformation and, the final regression loss is written as:

    Comparable with KLD,BD is also scale invariance.Therefore, the extended loss of BD also conforms to the previous three analysis.In subsequent experiments, it is proved that our loss has performance improvement in BD compared to the original function.

    3.4.Overall loss function

    We employ RetinaNet6as our baseline.The rotated bounding box is denoted as (x,y,w,h,θ).The overall training loss is:

    where hyper-parameter λ1,λ2control the trade-off.N indicates the number of anchors.The classification loss Lclsdenotes focal loss.6pndenotes the n-th probability of different classes,tnrepresents the label of n-th object.objnis a binary number,objn= 1 as objects and objn= 0 as background.Lα-Guassianis the regression loss where bnand gtnrepresent the n-th bounding box and ground-truth.

    4.Experiments

    We evaluate our proposed method on four typical datasets of remote sensing images, UCAS-AOD, HRSC2016, DIOR-R and DOTA.The following subsections introduce datasets,implementation details, ablation study and overall comparison, respectively.

    4.1.Datasets

    UCAS-AOD20consists of 1510 aerial images(659×1280 pixels) in two categories of 14596 objects, including vehicles and planes.In accord with DOTA,23we sample 1110 images as the training set and 400 as the testing at random.

    The HRSC201621dataset contains 1061 ship images which are captured from Google Earth and annotated with rotated box.The spatial resolution ranges from 2 m to 0.4 m.The sizes of images in the dataset cover from 300 × 300 to 1500 × 900.We split the dataset into training,validation and test set which contains 436 images, 181 images and 444 images respectively.

    DIOR-R22is an oriented object detection dataset composed of large-scale RS images, which contains 23463 images(800 × 800 pixels).The image resolution is between 0.5 m and 30 m.The dataset covers 20 common categories including 192518 instances.In our experiments, 11725 images are selected as training set, and the remaining 11738 images are used for testing.

    DOTA is one of the most commonly used public datasets in remote sensing consisting of 2806 images and the image size is 4000 × 4000 pixels.The 188282 instances in 15 categories are of different orientations, and labeled by an oriented quadrangle.The dataset is divided into three subsets, training set, validation set and test set for 1/2,1/6,1/3 of the whole dataset.All images in training and validation sets are segmented into subimages of 600×600 pixels,overlapping by 150 pixels,and the segmented images are rescaled to 800 × 800 as the training dataset for training.

    4.2.Implementation details

    We apply Tensorflow52and run all experiments with NVIDIA V100 GPUs.Experiments are initialized by Resnet-5053pretrained on ImageNet.54We use five pyramid layers {P3, P4,P5, P6, P7} to generate anchors.The batch size is set to 1 and the total epoch is 20 for all datasets.The weight decay and momentum of the momentum optimizer are set as 0.0001 and 0.9.and learning rate is 0.0005,and reduced at 8 and 14 epochs.For all datasets, we use the data augmentation for training,including random flipping and rotations.The image iterations per epoch for HRSC2016, UCAS-AOD, DOTA, and DIOR-R are 130000, 10000, 540000 and 282000 respectively.

    4.3.Ablation study

    We use RetinaNet as the base detector to conduct ablation study.Specifically, the baseline utilizes original FPN and smooth L1 regression loss unless noted specifically.

    4.3.1.Ablation test of DCFPN

    The results on four different datasets with DCFPN can be found in Table 1.The performance in four datasets, UCAS-AOD, HRSC2016, DIOR-R, and DOTA are increased by 0.46%, 3.15%, 3.34% and 1.30% respectively compared with the baseline (RetinaNet + FPN).The results show that the DCFPN is highly effective in detecting different kinds of objects.With a stronger R3Det detector, the highest performance improvement has also reached 2.36%in DIOR dataset.This also proves that when DCFPN is used to extract feature information, it is more advantageous to obtain the receptive fields of multi-scale objects in RS images.

    Table 1 Ablation studies of DCFPN on four different datasets.

    Fig.7 Ablation study of Power parameter α on DIOR-R dataset.Red dashed line denotes Lα-Guassian with α=0.5 while black dashed line denotes the baseline.

    Table 2 Ablation study of the hyperparameter on DIOR-R.

    4.3.2.Ablation test with different power parameter in α-Gaussian losses

    We analyze the availability of α-Gaussian losses, and the performance changes of α-Gaussian losses under different α values are compared on DIOR-R dataset.Here we set

    Table 4 Performance evaluation on horizontal detection.

    Table 3 Ablation studies of α-Gaussian across different regression losses on various datasets.‘α-Gaussian’ indicates that power transformation is used on regression loss.

    where α=0.5, τ=1, f(?)=sqrt, ε=1×10-3.As shown in Fig.7, it is evident that Lα-Guassianwith α ?(0.3,1) performs competitively well.When α=1 in baseline, the mAP is only 51.99.The performance reaches the optimal when α=0.5,with an increase of 3.57%.However, in extreme cases α=2,the convergence effect of model training is poor, and the mAP is only 41.30%,which is 10.69%lower than the baseline.Therefore, a proper choice of α is very important for performance.In our experiments, 0<α<1 is a better choice for Lα-Guassian.τ is a hyperparameter in the simple nonlinear transformation.As shown in Table 2,τ=1 performs the best in different cases.ε is regularization term and we set it to 1×10-3to make it effective.The results also prove that the α-Gaussian loss is a good choice for box regression.

    4.3.3.Ablation test with different α-Gaussian losses across several datasets

    We use four kinds of regression losses GWD, KLD, KFIoU and BD (mentioned in Section 3.3) on four datasets to prove the effectiveness and universality of α-Gaussian losses.α=0.5 is used for all these losses.The experimental results demonstrate that all four α-Gaussian losses outperform the original regression loss.Additionally, ‘α-Gaussian’ resulted in varying degrees of improvement on different kinds of regression losses on various datasets, as shown in Table 3.Lα-Guassian(KLD)shows the best performance among these loss functions.A greater performance improvement is obtained in the DIOR-R dataset of different losses, about 2.83%/0.18%/2.87%/1.65% on GWD/KLD/KFIoU/BD.This is probably because the fact that DIOR-R dataset has more categories and can comprehensively reflect the effect of the proposed loss transformation.

    With regard to the horizontal detection task,we have theoretically analyzed that the α-Gaussian losses can be used as the regression loss in the horizontal detection task in Section 3.3.To make the results more convincing,we continue to compare the smooth L1 loss for horizontal detection with our proposed α-Gaussian regression losses on the PASCAL VOC 2007,55a classic horizontal detection benchmark.We train on about 5 k images in VOC07 trainval set and get the results on VOC07 test set.As shown in Table 4, our proposed losses achieve comparable performance to the smooth L1 loss.In particular, Lα-Guassian(GWD) achieves an improvement of 2.96% compared with smooth L1.

    Table 6 Detection accuracy on DIOR-R.

    4.4.Overall comparison

    In the overall comparison, consisting with the state-of-the-art detection methods, we choose one-stage detector R3Det as base rotation detector to consider both speed and accuracy.Using the DCFPN structure proposed above,the loss function Lα-Guassian(KLD) which performs best in ablation study is selected to compare with other advanced rotation detectors with complex model structures.Specifically, data augmentation and multiscale training are used.

    4.4.1.Results on HRSC2016

    The HRSC2016 consists of large aspect ratio ships in arbitrary direction,posing a big challenge to the accurate localization in detection.As depicted in Table 5,7,10,13,16,18,28,29,32,42,44,56our method achieves relatively good performance, about 89.98%and 96.12% in term of VOC2007 and VOC2012 evaluation metric.Fig.8 shows the visual results of our method on HRSC2016 dataset.

    4.4.2.Results on DIOR-R

    DIOR-R is a large-scale dataset with massive categories and complex scenes.We compare our method against several advanced detectors on DIOR-R.We select various classes of objects at different scales, and scenes with dense or sparse object arrangements for visualization.The detection results are shown in Fig.9.Table 6 shows the specific performance of each category of objects.For the detection results of individual categories such as DAM,GC,TS and ST,there is quite some room to improve due to the small quantity of training instances per class, which is less than 1500.Likewise, some small object categories such as BR and VE have not achieved the optimal performance because the scale size is less than 80 pixels resulting difficulty in detecting accurately.Overall, our method outperforms most categories, and has achieved superior performance of about 55.56%.

    4.4.3.Results on DOTA

    We compare the performance of various detectors on the DOTA dataset.Our method can accurately obtain angle and position information through regression, as illustrated in Fig.10.Even in densely distributed scenarios such as large vehicles, this method can accurately detect arbitrary direction of objects without influencing the visual effect.From the experimental results of ground track field and ships,we notice that our method achieves superior detection ability for multiscale objects, and achieves accurate extraction of features at different scales.The detection results of current advanced methods in each category of DOTA are summarized in Table 7.We classify the detectors into one-stage and two-stage, and labeling the top two performances of each category severally.To reduce the impact of different training tricks, we mainly use Resnet-50(Res-50) as the backbone.It can be seen that we achieve the best performance among these detectors.At the same time, the results show a balanced and excellent performance for each category.Note that our proposed singlestage detector also achieves satisfactory results compared to the more complex feature extraction and processing of the two-stage detector.

    Table 7 Detection accuracy on DOTA.

    5.Conclusions

    To solve the problems of ineffective feature extraction and inaccurate rotated bounding box localization in remote sensing, we propose a new feature extraction structure called DCFPN, which utilizes dense multi-path dilation layers to cover all size ranges of objects in different scenarios.This structure can extract multi-scale information accurately and efficiently by generating larger and denser receptive fields,and perform feature transfer and fusion in deep networks.Meanwhile, a new transformation of losses is designed based on existing Gaussian modeling losses for more accurate bounding box regression in object detection.The properties are further analyzed to prove the effectiveness in theory.We conduct extensive ablation studies and experiments on four public datasets such as UCAS-AOD, HRSC2016, DIOR-R,and DOTA.Experimental results demonstrate that the proposed method outperforms the existing detectors on four challenging datasets, and demonstrates the effectiveness of dense context feature extraction and α-Gaussian loss on rotated object detection task.

    Declaration of Competing Interest

    The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

    在线av久久热| 在线永久观看黄色视频| 水蜜桃什么品种好| av国产精品久久久久影院| 久久ye,这里只有精品| 日韩大片免费观看网站| bbb黄色大片| 国产精品偷伦视频观看了| 香蕉丝袜av| 亚洲av电影在线进入| 国产欧美亚洲国产| 深夜精品福利| 一区二区日韩欧美中文字幕| 黑人巨大精品欧美一区二区蜜桃| 女同久久另类99精品国产91| 成人精品一区二区免费| 国产精品电影一区二区三区 | 日韩欧美免费精品| 黄色视频不卡| 精品少妇黑人巨大在线播放| av线在线观看网站| 老熟妇仑乱视频hdxx| 国产一区二区三区视频了| 美国免费a级毛片| 国产成人精品久久二区二区免费| 久久人人爽av亚洲精品天堂| 亚洲avbb在线观看| 夫妻午夜视频| 黄色片一级片一级黄色片| 久9热在线精品视频| 久久精品成人免费网站| 天堂中文最新版在线下载| 久久国产精品大桥未久av| 国产精品二区激情视频| 人妻一区二区av| 国产不卡av网站在线观看| 久久亚洲精品不卡| 日韩 欧美 亚洲 中文字幕| 免费黄频网站在线观看国产| 动漫黄色视频在线观看| 精品一区二区三区视频在线观看免费 | 俄罗斯特黄特色一大片| a级片在线免费高清观看视频| 色在线成人网| 亚洲国产欧美日韩在线播放| 欧美 日韩 精品 国产| 日韩欧美国产一区二区入口| 97在线人人人人妻| 欧美精品高潮呻吟av久久| 亚洲国产毛片av蜜桃av| 女人高潮潮喷娇喘18禁视频| 每晚都被弄得嗷嗷叫到高潮| 久久婷婷成人综合色麻豆| 中文亚洲av片在线观看爽 | 国产精品久久久久成人av| 国产成人av激情在线播放| 日日爽夜夜爽网站| 91老司机精品| 日日摸夜夜添夜夜添小说| 美女福利国产在线| netflix在线观看网站| 精品国产国语对白av| 一区二区三区乱码不卡18| 亚洲自偷自拍图片 自拍| 日韩免费高清中文字幕av| 在线播放国产精品三级| 99热网站在线观看| 亚洲综合色网址| 狂野欧美激情性xxxx| 亚洲综合色网址| 欧美大码av| 国产精品一区二区在线不卡| 黄色丝袜av网址大全| 91大片在线观看| 久久久精品94久久精品| 久久热在线av| 日本av手机在线免费观看| 美女福利国产在线| 91精品国产国语对白视频| 免费在线观看日本一区| 欧美精品av麻豆av| 人人澡人人妻人| 亚洲欧美日韩高清在线视频 | 国产视频一区二区在线看| 人妻久久中文字幕网| 99久久国产精品久久久| bbb黄色大片| 十分钟在线观看高清视频www| 日本wwww免费看| 欧美日本中文国产一区发布| 久久人人爽av亚洲精品天堂| 在线播放国产精品三级| 成人国语在线视频| 18禁裸乳无遮挡动漫免费视频| 美女高潮到喷水免费观看| 深夜精品福利| 成人免费观看视频高清| 黑丝袜美女国产一区| 少妇裸体淫交视频免费看高清 | 黄色怎么调成土黄色| 9色porny在线观看| 一级毛片精品| 午夜福利视频精品| 欧美 亚洲 国产 日韩一| 亚洲午夜精品一区,二区,三区| 亚洲欧美激情在线| 欧美亚洲日本最大视频资源| 亚洲少妇的诱惑av| 天天影视国产精品| 欧美激情 高清一区二区三区| 国产一卡二卡三卡精品| 91精品三级在线观看| 国产三级黄色录像| 91成年电影在线观看| 国产免费福利视频在线观看| 国产成人精品无人区| 99riav亚洲国产免费| 欧美日韩视频精品一区| 美女主播在线视频| 亚洲欧洲日产国产| 精品卡一卡二卡四卡免费| 国产精品熟女久久久久浪| 在线观看66精品国产| 老熟妇仑乱视频hdxx| 国产日韩欧美在线精品| 免费在线观看完整版高清| 午夜福利免费观看在线| 狠狠狠狠99中文字幕| 一区二区三区精品91| 欧美日韩视频精品一区| 女人精品久久久久毛片| 91大片在线观看| 亚洲午夜理论影院| 国产精品亚洲一级av第二区| 少妇粗大呻吟视频| 极品少妇高潮喷水抽搐| 国产一卡二卡三卡精品| 少妇被粗大的猛进出69影院| 亚洲久久久国产精品| 久久久水蜜桃国产精品网| 久久中文字幕一级| 日本av免费视频播放| 最新在线观看一区二区三区| 中文字幕最新亚洲高清| 大片免费播放器 马上看| 亚洲成人免费电影在线观看| 男人舔女人的私密视频| 国产精品久久久人人做人人爽| 国产免费现黄频在线看| a在线观看视频网站| 成年女人毛片免费观看观看9 | 日韩有码中文字幕| 日韩中文字幕视频在线看片| 国产成人一区二区三区免费视频网站| 法律面前人人平等表现在哪些方面| 久久 成人 亚洲| 亚洲色图av天堂| 亚洲国产成人一精品久久久| 人妻 亚洲 视频| 老司机影院毛片| 久久中文字幕一级| 一级,二级,三级黄色视频| 无人区码免费观看不卡 | 国产精品麻豆人妻色哟哟久久| 精品乱码久久久久久99久播| 精品国产乱码久久久久久男人| 欧美乱妇无乱码| 精品一区二区三区四区五区乱码| 国产精品一区二区免费欧美| 色婷婷久久久亚洲欧美| 亚洲精品中文字幕在线视频| 欧美精品高潮呻吟av久久| 十八禁网站免费在线| 51午夜福利影视在线观看| 中文字幕人妻熟女乱码| 18禁裸乳无遮挡动漫免费视频| 高潮久久久久久久久久久不卡| 757午夜福利合集在线观看| 婷婷成人精品国产| 好男人电影高清在线观看| 久久精品国产亚洲av香蕉五月 | 操出白浆在线播放| 国产成人av教育| 色尼玛亚洲综合影院| 黑人操中国人逼视频| 午夜视频精品福利| 日韩欧美一区视频在线观看| 黑丝袜美女国产一区| 后天国语完整版免费观看| 大码成人一级视频| 脱女人内裤的视频| 丰满饥渴人妻一区二区三| 亚洲第一av免费看| 国产精品国产av在线观看| a级片在线免费高清观看视频| 啦啦啦 在线观看视频| 一级片免费观看大全| 日韩一卡2卡3卡4卡2021年| 国产精品欧美亚洲77777| 美女福利国产在线| 日韩中文字幕欧美一区二区| 亚洲美女黄片视频| 深夜精品福利| 夜夜爽天天搞| 国产成人av教育| 怎么达到女性高潮| 久久国产精品影院| 亚洲欧美色中文字幕在线| 国产一区二区三区综合在线观看| 亚洲av日韩在线播放| 亚洲国产av新网站| 高潮久久久久久久久久久不卡| 一级黄色大片毛片| 亚洲精品在线观看二区| 欧美老熟妇乱子伦牲交| 亚洲精品中文字幕一二三四区 | 久久久久网色| 国产亚洲av高清不卡| 亚洲 国产 在线| 午夜福利,免费看| 91av网站免费观看| 中文字幕人妻熟女乱码| svipshipincom国产片| av国产精品久久久久影院| 成年版毛片免费区| 天天躁夜夜躁狠狠躁躁| 精品乱码久久久久久99久播| 国产成人精品无人区| 日本vs欧美在线观看视频| 50天的宝宝边吃奶边哭怎么回事| 老熟妇乱子伦视频在线观看| 在线观看www视频免费| 国产男靠女视频免费网站| 黄色视频不卡| 久久九九热精品免费| 777米奇影视久久| 啦啦啦在线免费观看视频4| 黄网站色视频无遮挡免费观看| 国产不卡一卡二| 一个人免费在线观看的高清视频| 在线观看免费日韩欧美大片| 久久ye,这里只有精品| 又黄又粗又硬又大视频| 99国产极品粉嫩在线观看| www.999成人在线观看| 电影成人av| 久久久精品国产亚洲av高清涩受| 午夜免费鲁丝| 久久久精品区二区三区| 久久精品aⅴ一区二区三区四区| 性色av乱码一区二区三区2| 国产国语露脸激情在线看| 黄色a级毛片大全视频| 亚洲中文字幕日韩| 久久亚洲精品不卡| 亚洲av第一区精品v没综合| 亚洲熟女精品中文字幕| 黄色丝袜av网址大全| 欧美成人午夜精品| 国产91精品成人一区二区三区 | 岛国在线观看网站| 国产精品麻豆人妻色哟哟久久| 在线看a的网站| 中文字幕色久视频| 久久精品亚洲熟妇少妇任你| 亚洲第一av免费看| 国产精品av久久久久免费| 亚洲综合色网址| 激情在线观看视频在线高清 | 丝袜喷水一区| 国产成人精品在线电影| 国产一区二区在线观看av| 亚洲专区字幕在线| 乱人伦中国视频| 亚洲熟妇熟女久久| av福利片在线| 国产免费视频播放在线视频| 欧美激情久久久久久爽电影 | 日韩欧美三级三区| 欧美成狂野欧美在线观看| 中文字幕av电影在线播放| 成人国产一区最新在线观看| 精品国产一区二区三区久久久樱花| 午夜老司机福利片| 在线观看舔阴道视频| 男女午夜视频在线观看| 免费女性裸体啪啪无遮挡网站| 可以免费在线观看a视频的电影网站| 无限看片的www在线观看| 免费少妇av软件| 久久免费观看电影| 黄片播放在线免费| 悠悠久久av| 成人免费观看视频高清| 日韩大片免费观看网站| 久久人人爽av亚洲精品天堂| 丝袜在线中文字幕| 日韩视频一区二区在线观看| 久久精品国产99精品国产亚洲性色 | 免费人妻精品一区二区三区视频| 成人18禁在线播放| 国产亚洲av高清不卡| 日韩欧美三级三区| 18禁黄网站禁片午夜丰满| 国产成人精品久久二区二区免费| 91成人精品电影| 亚洲一码二码三码区别大吗| 亚洲精品国产色婷婷电影| 免费在线观看影片大全网站| 成人亚洲精品一区在线观看| 久久精品国产99精品国产亚洲性色 | 久久人妻熟女aⅴ| 91麻豆av在线| 欧美日韩国产mv在线观看视频| 欧美在线黄色| 亚洲av电影在线进入| 老司机午夜十八禁免费视频| 天堂8中文在线网| 午夜精品国产一区二区电影| 国产精品自产拍在线观看55亚洲 | 这个男人来自地球电影免费观看| 久久久精品94久久精品| 一二三四在线观看免费中文在| 美女高潮喷水抽搐中文字幕| 99热国产这里只有精品6| 高清欧美精品videossex| 久久久精品免费免费高清| 如日韩欧美国产精品一区二区三区| 国产99久久九九免费精品| 久久精品国产a三级三级三级| 性色av乱码一区二区三区2| 国产亚洲午夜精品一区二区久久| 日韩中文字幕欧美一区二区| 美女高潮到喷水免费观看| 人人妻人人澡人人看| 成年版毛片免费区| 欧美日韩中文字幕国产精品一区二区三区 | 欧美在线黄色| 一边摸一边抽搐一进一小说 | 91老司机精品| 自线自在国产av| 亚洲七黄色美女视频| 久久ye,这里只有精品| 变态另类成人亚洲欧美熟女 | 精品国产一区二区三区久久久樱花| 久久久久国产一级毛片高清牌| 久久久久网色| 免费在线观看完整版高清| 久久这里只有精品19| 一边摸一边抽搐一进一出视频| 美女高潮喷水抽搐中文字幕| 麻豆国产av国片精品| 日韩免费高清中文字幕av| 日韩精品免费视频一区二区三区| 丝袜人妻中文字幕| 精品国产乱子伦一区二区三区| 91字幕亚洲| 国内毛片毛片毛片毛片毛片| 日韩中文字幕视频在线看片| 一边摸一边做爽爽视频免费| 视频区欧美日本亚洲| 亚洲avbb在线观看| 在线观看免费视频日本深夜| 国产精品九九99| 久久精品国产99精品国产亚洲性色 | 国内毛片毛片毛片毛片毛片| 夫妻午夜视频| 婷婷成人精品国产| 欧美日韩亚洲综合一区二区三区_| 国产精品欧美亚洲77777| 久久婷婷成人综合色麻豆| 免费观看av网站的网址| 大片电影免费在线观看免费| 国产视频一区二区在线看| 亚洲久久久国产精品| 国产成人免费观看mmmm| 一边摸一边做爽爽视频免费| 国产午夜精品久久久久久| av网站在线播放免费| 超碰97精品在线观看| 国产有黄有色有爽视频| 高清av免费在线| 久久久精品区二区三区| 丝袜美腿诱惑在线| 最近最新中文字幕大全电影3 | 日韩欧美一区视频在线观看| 又大又爽又粗| 亚洲成人手机| 精品国产一区二区三区四区第35| 国产精品成人在线| 欧美乱码精品一区二区三区| 久久久精品免费免费高清| 日韩有码中文字幕| 欧美国产精品va在线观看不卡| 国产av一区二区精品久久| 两性夫妻黄色片| 国产av国产精品国产| 精品一区二区三卡| 波多野结衣一区麻豆| 50天的宝宝边吃奶边哭怎么回事| 男女午夜视频在线观看| 国产精品成人在线| 一本大道久久a久久精品| 1024香蕉在线观看| 老司机午夜十八禁免费视频| 国产精品免费视频内射| 国产男女超爽视频在线观看| 国产片内射在线| 精品人妻1区二区| 成人手机av| 一本久久精品| 亚洲精品美女久久久久99蜜臀| 国产成人免费观看mmmm| 又紧又爽又黄一区二区| 久久久国产精品麻豆| 亚洲欧洲日产国产| 午夜免费成人在线视频| 母亲3免费完整高清在线观看| 中文字幕高清在线视频| avwww免费| 国产精品美女特级片免费视频播放器 | 久久精品国产综合久久久| 国产日韩欧美在线精品| 麻豆av在线久日| 精品高清国产在线一区| 热99久久久久精品小说推荐| aaaaa片日本免费| 国产精品99久久99久久久不卡| 国产精品.久久久| 国产人伦9x9x在线观看| 在线观看一区二区三区激情| 一区在线观看完整版| 国产精品一区二区精品视频观看| 一个人免费在线观看的高清视频| 91老司机精品| 精品一区二区三卡| 操出白浆在线播放| 亚洲熟女毛片儿| 中文字幕人妻丝袜制服| 成人18禁在线播放| 国产亚洲精品第一综合不卡| 免费观看a级毛片全部| 在线观看66精品国产| 肉色欧美久久久久久久蜜桃| 亚洲色图综合在线观看| 一区福利在线观看| 亚洲国产av影院在线观看| 动漫黄色视频在线观看| 在线看a的网站| 激情视频va一区二区三区| 亚洲avbb在线观看| 老熟妇仑乱视频hdxx| 亚洲av国产av综合av卡| 国产成人精品久久二区二区免费| 亚洲av美国av| 久久久久国产一级毛片高清牌| 丝袜在线中文字幕| 久久免费观看电影| 午夜福利影视在线免费观看| 一进一出好大好爽视频| 在线av久久热| 欧美亚洲日本最大视频资源| 精品福利永久在线观看| 精品久久久精品久久久| 亚洲视频免费观看视频| 精品欧美一区二区三区在线| 成人国产一区最新在线观看| 亚洲五月婷婷丁香| 777久久人妻少妇嫩草av网站| 香蕉丝袜av| 精品少妇内射三级| 国产精品免费一区二区三区在线 | 国产精品久久久久久人妻精品电影 | 另类精品久久| 亚洲av欧美aⅴ国产| 天天躁日日躁夜夜躁夜夜| 一二三四社区在线视频社区8| 国产单亲对白刺激| 国产成人欧美| 嫁个100分男人电影在线观看| 欧美日韩av久久| 亚洲一码二码三码区别大吗| 免费在线观看日本一区| 人妻 亚洲 视频| 十分钟在线观看高清视频www| 一边摸一边做爽爽视频免费| 亚洲一码二码三码区别大吗| 成人永久免费在线观看视频 | 怎么达到女性高潮| 黄片小视频在线播放| 久久久久国产一级毛片高清牌| 99国产极品粉嫩在线观看| 日韩有码中文字幕| 丰满饥渴人妻一区二区三| 亚洲国产欧美一区二区综合| 天天躁日日躁夜夜躁夜夜| 久久久国产精品麻豆| 久久午夜综合久久蜜桃| av欧美777| 久久久国产一区二区| 一个人免费看片子| 后天国语完整版免费观看| 欧美成人午夜精品| 中亚洲国语对白在线视频| 99国产精品一区二区三区| 叶爱在线成人免费视频播放| 久久精品国产综合久久久| 性高湖久久久久久久久免费观看| 成人国语在线视频| 乱人伦中国视频| 日本一区二区免费在线视频| 亚洲精品中文字幕在线视频| 久久精品成人免费网站| 成人精品一区二区免费| 啦啦啦 在线观看视频| 亚洲精品久久午夜乱码| 十八禁网站网址无遮挡| 91大片在线观看| 在线观看一区二区三区激情| 成人国产av品久久久| 女人久久www免费人成看片| 日本av手机在线免费观看| 欧美黄色片欧美黄色片| 欧美大码av| 少妇被粗大的猛进出69影院| 免费久久久久久久精品成人欧美视频| tube8黄色片| 可以免费在线观看a视频的电影网站| 涩涩av久久男人的天堂| 老司机午夜十八禁免费视频| 丝瓜视频免费看黄片| 精品熟女少妇八av免费久了| 王馨瑶露胸无遮挡在线观看| 国产三级黄色录像| 真人做人爱边吃奶动态| 成人免费观看视频高清| 丁香六月欧美| 亚洲精品粉嫩美女一区| 夫妻午夜视频| 考比视频在线观看| 国产精品 欧美亚洲| 日韩精品免费视频一区二区三区| 老司机福利观看| 在线观看免费视频网站a站| 久久久国产一区二区| 男女无遮挡免费网站观看| 亚洲综合色网址| 一区福利在线观看| 多毛熟女@视频| 另类亚洲欧美激情| 正在播放国产对白刺激| 1024香蕉在线观看| 国产成人av激情在线播放| 久久av网站| 午夜激情久久久久久久| 精品国产一区二区三区四区第35| 午夜精品久久久久久毛片777| 欧美日韩亚洲高清精品| 日本av免费视频播放| 欧美亚洲日本最大视频资源| 操出白浆在线播放| 日韩欧美国产一区二区入口| 最新在线观看一区二区三区| 在线观看一区二区三区激情| 欧美中文综合在线视频| 亚洲av国产av综合av卡| 欧美亚洲日本最大视频资源| 久久香蕉激情| 亚洲精品国产区一区二| 波多野结衣一区麻豆| 久久久久久久国产电影| 一区二区av电影网| 亚洲国产毛片av蜜桃av| 精品欧美一区二区三区在线| 国产又爽黄色视频| 国产精品 欧美亚洲| 国产91精品成人一区二区三区 | 亚洲欧美色中文字幕在线| 国产男女内射视频| 日韩视频在线欧美| 人妻久久中文字幕网| 欧美黑人精品巨大| 免费久久久久久久精品成人欧美视频| 人人妻,人人澡人人爽秒播| 三级毛片av免费| 丁香六月天网| 欧美成人免费av一区二区三区 | 热re99久久国产66热| 老汉色av国产亚洲站长工具| 777米奇影视久久| 精品国产一区二区三区四区第35| 国产精品免费大片| 国产不卡av网站在线观看| 午夜免费鲁丝| 欧美激情极品国产一区二区三区| 色综合婷婷激情| 国产精品av久久久久免费| 成年人午夜在线观看视频| 久久99一区二区三区| 女同久久另类99精品国产91| 国产精品成人在线| 中文字幕制服av| 日本撒尿小便嘘嘘汇集6| 国产成人系列免费观看| 午夜成年电影在线免费观看| 欧美激情 高清一区二区三区| 免费在线观看日本一区| 女人久久www免费人成看片| 久久国产亚洲av麻豆专区| 精品国产乱码久久久久久男人| 久久久久网色| 大香蕉久久成人网| www.自偷自拍.com| 自拍欧美九色日韩亚洲蝌蚪91| 成年动漫av网址|