Jiarui Liang 1, Tingbo Zhang 2, Xiaolin Tian 1,3,*
1Faculty of Innovation Engineering, Macau University of Science and Technology, Macau, China;
2School of Mechanical & Aerospace Engineering, Nanyang Technological University, Singapore;
3The State Key Laboratory of Lunar and Planetary Sciences, Macau University of Science and Technology, Macau, China.
*Corresponding Author’s Email: xltian@must.edu.mo
Abstract: The impact crater is one of the most significant and conspicuous geologic features on the Lunar. The study of Lunar impact crater detection is crucial in spacecraft Lunar landings and Lunar surface resources development. Currently, the YOLO network is widely used in the field of object detection. However, few articles discuss which modifications can improve the experimental results of YOLO network in detecting lunar impact craters, as well as the detection effects when several modifications are made to YOLO network. This paper investigates the effects of adding the Dynamic Head Attention Mechanism (DyHead), Deformable Convolutional Network (DCN), and Wise-IoU (WIoU) loss function when using YOLOv7. It also explores the effects of using these modifications individually and in combination. We have conducted several experiments to evaluate the performance of the newly proposed model, compared with the original YOLOV7, the improved structure achieves better results. This can provide more reliable assistance for researchers’ detection work.
Keywords: Crater; Yolo Network; Deformable Convolutional Networks; Dynamic Head Attention Mechanism; Deep Learning;
- Introduction
The impact crater is one of the most prominent geological features on the Lunar surface. These craters are formed by meteorite or asteroid impacts and typically exhibit a bowl-shaped appearance, characterized by a relatively flat floor encircled by elevated rim walls. The study of the Lunar impact craters began early in history. These comprehensive studies of the Lunar impact craters provide an adequate basis for Lunar landings and the exploitation of Lunar surface resources [1–2]. The Lunar impact craters are classified into five distinct categories based on their morphological characteristics: Albategnius C, Biot, Sosigenes, Triesneker, and Tycho [3]. The high-resolution remote sensing satellite images of the Lunar have been obtained by orbital cameras. These mainly include Charge- Coupled Devices (CCD) data and Digital Elevation Model (DEM) data. Based on these global Lunar remote sensing satellite images, various databases of Lunar impact craters have been progressively established using both manual and automated labeling techniques [4-5]. Then the research on Lunar impact crater detection was conducted based on the reliable crater database. However, these impact craters on the Lunar are not static, their shapes and sizes dynamically change due to various geological activities, including space weathering and moonquakes. Researchers study the ages of Lunar impact craters by analyzing changes in their morphological features, providing a scientific basis for determining the relative stratigraphic ages of newly observed impact craters [6], and updating the Lunar impact crater databases. According to the pre-established databases, we can obtain the coordinate information of the studied craters and apply them in further detection experiments.
Initial approaches to crater detection employed traditional image-processing techniques, including Morphology-Based Algorithms, Object-Oriented Flooding Methods, Template Matching Strategies, Edge Detection, and Hough Transform [7-8]. However, the complex and varied topography of the Lunar surface, fluctuating lighting conditions, and other environmental factors often exceed the capabilities of these algorithms, which require manual intervention to address exclusive or anomalous scenarios. These tasks are usually time-consuming and error-prone. Therefore, researchers have gradually begun to adopt machine learning techniques like Supported Vector Machine or Random Forest to efficiently and accurately detect craters detect craters [9]. In 2012, a new Convolutional Neural Network (CNN) named “AlexNet” achieved breakthrough results in the ILSVRC- 2012 competition, significantly reducing the error rate in image classification. Marking a revolutionary milestone for deep learning and demonstrating its immense potential in computer vision and image processing. In the object detection area, deep learning solutions like Faster R-CNN, YOLO, SSD, RetinaNet, and EfficientDet are occupying the dominant position because of their powerful automatic feature learning ability and real-time performance [10]. Consequently, the use of CNN techniques for the detection of Lunar impact craters has been increasingly embraced by researchers.
Based on the Lunar remote sensing image data from Lunar Reconnaissance Orbiter Camera (LROC) Narrow Angle Camera (NAC) , Chen et al. [11] utilized an improved Faster R-CNN to identify and estimate the age of the small Lunar crater, demonstrating the effectiveness of the CNN models on detecting the small craters . Jiang et al. [12] developed a multitype landform detector using the lightweight SSD model that achieved notable resilience in landform detection in different environments. Despite these deep learning models demonstrating superior performance in detection tasks, the YOLO is particularly favored for Lunar impact crater detection tasks because of its rapid speed and low false positive rates. Each version of the YOLO series has been studied and applied in this research orientation. Qiu et al. [13] constructed a new small-scale CNN model based on the YOLOv5 that enhanced the processing speed and accuracy. Through literature research, we found that YOLO achieves good results in detecting small and medium-sized objects in most complex scenarios [14-15]. Therefore, we believe that making targeted improvements based on the YOLO network structure and using it to identify lunar impact craters is an interesting experimental topic.
In this paper, we proposed a novel model, Dynamic YOLOv7, for the mission of Lunar impact crater detection and reached both lower processing time and higher detection accuracy. The Dynamic YOLOv7 is improved from YOLOv7 by substituting the convolutional modules in the backbone with the deformable convolution of Deformable Convolutional Networks version 2 (DCNv2), incorporating Dynamic Head attention mechanism into the inference procedure, and replacing the default bounding-box loss function into the WIoU. Control variable experiments have been made on the CCD crater dataset to evaluate the performance of the Dynamic YOLOv7.
- Dataset and Methodology
2.1 DataSet Preparation
The data utilized in this paper is processed from the CCD data “Moon LRO LROC WAC Global Morphology Mosaic 100m” published by USGS Astrogeology Science Center, see Figure 1. The dataset applied in this project contains 240,000 labeled targets. The images for training have 190,000 labeled targets, and 50,000 for validation.
Figure. 1 Moon LRO LROC WAC Global Morphology Mosaic 100m.
Based on the Robinson crater database [4], which includes latitude, longitude, and radius information for Lunar impact craters with diameters of 1 kilometer or more, we utilized the Platte-Carree projection technique to correct the distortions of these cratersfrom the CCD data with reference coordinates. By matching the craters on the cropped images with the positional information from the database, we created the corresponding label files. The cropped and labeled Lunar images are illustrated in Figure 2.
Figure 2. Cropped and Labeled Lunar data.
2.2 Methodology design
2.21 YOLOv7 Network Architecture
The YOLOv7 [16] is an outstanding representative in the field of single-stage object detection models with high accuracy and fast speed. It is a popular object detection method that used across a variety of computer vision systems, and its excellent performance makes it an ideal model for detecting Lunar impact craters. The YOLOv7 consists of backbone, neck, and head. Its architecture is illustrated in Figure 3. The backbone is responsible for extracting useful information from the input image. This information will later be processed by the detection heads. In the YOLOv7 network architecture, the backbone is composed of ELAN, MP, and CBS modules. The CBS module is composed of three components: convolutional layer, batch normalization layer, and SiLU activation function layer. The MP module has two branches. The first branch is composed of a max pooling layer for Down Sampling and a convolutional layer to change the number of channels. The other branch is composed of two convolutional layers. These two branches are concatenated together to implement super Down Sampling. The ELAN module is a feature aggregation strategy that enables the network to extract more features by controlling the gradient paths. It is composed of two branches as well. The first branch is a convolutional layer and the second branch is a channel transformation convolutional layer and feature extraction convolutional layers. The neck contains an additional module compared with the backbone, which is SPPCSPC. The SPPCSPC is designed to integrate the benefits of Spatial Pyramid Pooling (SPP) and Cross Stage Partial (CSP) structures. The SPP enhances the receptive field by applying maximum pooling on multiple scales. In the CSP part, the feature map is divided into two parts, one part is processed conventionally, the other part is processed by SPP sequence, and then the two parts are concatenated. The head receives and processes the feature extracted from previous sectors and employs anchor boxes to output feature maps, and go through the detection and inference process to generate the final output of this model. It contains the REP module and the CBM module. The REP module is first processed with a convolution layer and batch normalization layer, and then divided into two paths: one through a 3×3 convolution layer and batch normalization layer, and another through a 1×1 convolution layer and batch normalization layer, and finally the output of these two paths is concatenated with the first input. The CBM module is similar to the CBS module with the SiLU activation function being replaced by the sigmoid activation function. The structures of these modules included in the YOLOv7 network architecture are illustrated in Figure 4.
Figure 3. YOLOv7 network architecture.
Figure 4. Components of the modules in YOLOv7.
2.22 Deformable Convolutional Networks
The DCN [17] is proposed as an improved CNN that can handle geometric transformations within the target by augmenting spatial location sampling in the modules with additional offsets and offsets-based learning. The traditional CNN utilized in YOLOv7 lacks of adaptability, and the sampling locations are usually fixed. By replacing part of the CBS modules in the YOLOv7 network architecture with the DCN, the capabilities of YOLOv7 are improved when facing the more complex situation of handling deformable transformations. This is essential in the mission of Lunar impact crater detection as the craters in the image may be stretched or distorted due to different shooting angles and post-image processing. The formula of the traditional convolution is represented as:
In formula (1), the output y(p0) at a position p0 is calculated as the sum of the weighted inputs x(p0 + pn) across all positions pn within a region R. Each input is multiplied by a corresponding weight w(pn), and then all these products are summed together. In order to introduce the deformable convolution based on the traditional convolution, it is necessary to augment the regular grid R with additional offsets. This can be represented as:
The set of additional offsets with specifies the adjustments for each sampling location. In real situations, the offsets are usually fractional numbers rather than integers, diverging from the integral coordinates of the input feature map.So we need to apply the bilinear interpolation method to retrieve the values of the offset positions. The bilinear interpolation method is depicted as:
In formula (3), x(p) represents the output at position p, and the G(q, p) is a bilinear interpolation kernel, which applies weights to the inputs based on their spatial relationship to the location of x(p). The x(q) denotes the input values at position q, which are integrated spatial locations in the input feature map.
Based on the formulas above, we have obtained the formula of the deformable convolution of DCN. It is represented as follows:
Based on the DCN, the DCNv2 is proposed with more deformable and better results. The DCNv2 introduces a novel modulation mechanism within its deformable convolutionmodule. This mechanism allows for adjustment in the sampling position offsets and modulation of feature amplitude at these positions. According to the formula (5), the deformable convolution of DCNv2 is depicted as:
The is the newly introduced modulation scalar at position pn. The architecture of the deformable convolution is illustrated in Figure 5.
In this paper, we modified several ELAN and MP modules in the YOLOv7 network architecture by replacing the CBS blocks with the DCN blocks. The DCN block represents the deformable convolution of the DCNv2. We have proposed two new modules, D-ELAN and MP-D with the DCN block modification. In the D-ELAN module, the CBS on the second branch is replaced by DCN, and the last three blocks of the CBS on the first branch are all replaced by DCN blocks. In the MP-D module, only the last block of the CBS on the second branch is replaced by DCN. The architecture of the new modules illustrated in Figure 6.
Figure 5. Architecture of the deformable convolution of DCNv2.
Figure 6. Architecture of the D-ELAN and MP-D.
3.2.3. Dynamic Head Attention Mechanism
The DyHead attention mechanism is an advanced technique in CNN that enhances model perception across different scales, spatial locations, and tasks. This method goes beyond traditional attention mechanisms by addressing all three dimensions—scale, spatial, and task awareness simultaneously within a unified framework. By standardizing and reformatting feature maps at various levels of the feature pyramid, DyHead can handle multiple attention demands efficiently. This results in significantly improved performance and flexibility in target detection tasks, as it dynamically adjusts sampling positions and feature amplitudes. To interpret the DyHead attention mechanism, we first represent a concatenation of features , where L denotes the number of levels in a feature pyramid. These features are resized by up sampling or down sampling to match the scale across pyramid levels. The resized feature pyramid forms a four-dimensional tensor , with H as height, W as width, and C as the channel count at the median pyramid level. The feature tensor can be represented in three dimensions as , where S = H × W. Based on the predefined representation, the DyHead an be expressed. For a general attention mechanism applied to a feature tensor, it is represented as , where π(·) denotes the attention mechanism function that contains scale-aware, spatial-aware, and task-aware attention.
For scale-aware attention, it is represented as:
Using a 11 convolutional layer, the function is approximated to operate on the feature tensor F . The resulting features are then activated by a hard sigmoid function, This activation function allows the integration of features extracted at different scales according to their semantic features, effectively enabling the model to handle multi-scale feature extraction and enhancing its capability to process varying feature sizes within the input data.
For scale-aware attention, it is represented as:
In formula (7), corresponds to a component similar to that in DCN. Specifically, denotes the original anchor position, while represents the learned offset relative to this position. The variable k indicates the number of sparse sampling locations, and is the self-learned importance scalar at . Based on the central level of the feature set F, the spatial offset and importance factor can be derived. This formulation aligns with the principle of deformable convolution, which introduces sparsity in attention earning and subsequently aggregates features at identical spatial locations across different levels.
For task-aware attention, it is represented as:
This component is designed to selectively emphasize certain channels Fc within the feature tensor F. The Fc denotes the feature slice at the c-th channel of the tensor F. α1(F) and α2(F) are learnable coefficients dependent on the feature tensor F, used to scale the feature slice Fc. They are dynamically adjusted through a hyper-function. β1(F) and β2(F) are learnable bias terms dependent on the feature tensor F, used to adjust the scaled feature values. The max operation selects the higher value, optimizing the feature activation.
Based on the three dimension of the awareness, the DyHead integrated them together
as:
In the newly proposed model, the DyHead is introduced into the IDetect procedure of the Dynamic YOLOv7. The IDetect is responsible for converting the output feature map of the network into detection results. The IDetect contains the procedure of input feature map, convolution operations, anchor boxes, prediction generation, activation functions, and post-processing. In the Dynamic YOLOv7, the DyHead sequence processing is added after the input feature map procedure. This operation enables the DyHead module to refine the input features at multiple levels, resulting in better detection and recognition of objects in different contexts and improved overall detection performance.
3.2.4. WIoU Loss Function
The default bounding-box loss function utilized in YOLOv7 is CIoU [18]. CIoU enhances the standard IoU by considering the overlapping area, center point distance, and aspect ratio between the detection frame and the real frame. However, CIoU applies the same static weights to all samples, potentially leading to overfitting high-quality samples while neglecting common and low-quality samples. In contrast, WIoU introduces a dynamic non-monotonic focusing mechanism that intelligently adjusts weights based on sample quality. This approach enhances the model’s ability to generalize across different sample qualities, thereby improving overall detection accuracy. The WIoU not only focus on the intersection and concurrency ratio of the anchor frames to the target frames, but also dynamically adjusts the gradient gain of the loss function based on the quality of the anchor frames.
The WIoU introduces an exponential function RWIoU that adjusts for the distance between the center point of the anchor box and the ground truth box. The exponential function is depicted as:
The is the square of the Euclidean distance between the center point of the anchor frame (x, y) and the center point of the target frame . This distance measures the degree of offset between the two boxes. TheWg and Hg represent the width and height of the smallest enclosing box, respectively. The minimum surrounding box is the smallest rectangle that contains both the anchor box and the target box. There presents the area of the minimum boxes. The (·)∗ indicates that the two values are separated from the computation graph to prevent them from creating gradients in back propagation. The first version of the WIoU is established with the exponential
function, it is represented as:
Based on the first version of the WIoU, the WIoUv2 introduces the monotonic focusing coefficient to improve the focusing capability of high-quality anchor frames, optimized with fixed gradient gain. It is represented as:
The focusing factor is used to reduce the weight of easy-to-learn samples (smaller loss values) and amplify the weight of difficult-to-learn samples. By adjusting dynamically in this way, the model is able to pay more attention to the difficult-to-learn samples during the training process, improving the performance and robustness of the model.
Compared to WIoUv2, WIoUv2 is an improvement on WIoUv1 in another orientation. It improves the generalization ability of the model by introducing a non-monotonic focusing coefficient which allows the model to dynamically adjust the gradient gain according to the real-time changes of the anchor frame quality. It is represented as:
The β denotes the outlier degree, α and δ are hyperparameters used to control the gradient gain.
In this paper, we finally applied the WIoUv3 to replace the CIoU bounding box loss function in the YOLOv7 after experiment to empower the new model Dynamic YOLOv7 with more flexibility and robustness.
Based on the materials discussed above, the network architecture of the Dynamic
YOLOv7 is illustrated in Figure 7.
Figure 7. Architecture of the Dynamic YOLOv7.
- Results and discussion
2.1Experiments and Results
3.11 Experimental Environment Deployment
Details of the experiment and the hyperparameter settings are given as follows:
- Cloud Platform: AutoDL;
- GPU: RTX 4090(24GB) * 1;
- Python: 3.8.10;
- Pytorch: 1.13.1;
- CUDA: 12.2;
- Batch Size: 8;
- Optimizer: SGD;
- Learning Rate: 0.01.
3.22 Evaluation Metrics
To evaluate the performance of the newly proposed model comprehensively, we have applied the metrics of Precision, Recall, mAP@0.5, and mAP@0.5:0.95. The Precision measures the proportion of true positive predictions among all positive predictions made by the model which is illustrated in formula (14). The Recall measures the proportion of true positive predictions among all actual positive cases which is illustrated in formula (15). The mAP@0.5 is the mean Average Precision when the IoU threshold is 0.5 which is illustrated in formula (16). And mAP@0.5:0.95 is the mean Average Precision when IoU thresholds range from 0.5 to 0.95 which is illustrated in formula (17).
The TP indicates the number of correct positive predictions, FP indicates number of incorrect positive predictions, and the FN indicates the number of actual positive cases that were incorrectly predicted as negative.
3.23 Results
In this paper, we allocated the dataset into two distinct subsets for the training process: a training set and a validation set. We divided the data randomly according to a 8:2split ratio, where 80% of the data was used for training purposes, and 20% for validation. This approach ensures that the model is trained on a substantial portion of the data while still maintaing a significant subset for evaluating its performance and generalization capabilities. To validate the efficacy of each improvement made to the YOLOv7 model, we conducted a series of ablation experiments using the proposed dataset. Initially, we trained the baseline YOLOv7 model to establish a reference point. Subsequently, we incrementally introduced additional components to assess their impact. The first improvement involved integrating the WIoU module. After that, we added the DCNv2 module. Finally, we combined all three components, including the DyHead, to assess the combined impact of these improvements on the performance of the model. The ablation experiment results are illustrated in Table 1. Partial algorithm detection results are shown in Figure 8 and Figure 9.
Table 1. Ablation Experiments Results.
| Model | GROUP | WIoU | DCNv2 | DyHead | Precision | Recall | mAP@0.5 | mAP@0.5:0.95 |
| YOLOv7 | 1 | 76.60% | 70.31% | 80.51% | 50.85% | |||
| 2 | ✓ | 77.91% | 69,89% | 80.57% | 50.87% | |||
| 3 | ✓ | ✓ | 78.75% | 71.74% | 81.38% | 51.33% | ||
| 4 | ✓ | ✓ | ✓ | 79.83% | 72.21% | 81.79% | 51.59% |
Figure 8. Detection results of the Dynamic YOLOv7
Figure 9. Detection results of the YOLOv7.
3.3 Comparison between Improved structure and Original YOLOv7
The comparison results indicate that Dynamic YOLOv7 exhibits bettrer performance when compared with YOLO v7. As shown in Table 1, this table presents the results of the ablation experiments. Compared to YOLO v7, our model achieves an precosion rate improvement from 76.60% to 79.83%, and a recall rate improvement from 70.31% to 72.21%. mAP@0.5 improved from 80.51%to81.79%, and mAP%0.5:0.95 improved from 50.85% to 51.59%. This indicates that our new algorithm outperforms the original YOLO v7 in terms of both crater detection capability and false positive rates.
From the comparison of Figures 8 and 9, we find that the new algorithm has a lower error rate than the original YOLOv7 model when detecting small and medium-sized impact craters.
- Conclusion
This paper presents an experimental study on whether improvements to the YOLOv7 architecture can enhance the ability to detect lunar impact crater. The improvements include adding a dynamic head attention mechanism (DyHead), using a deformable convolutional network (DCN), and replacing the loss function with WIoU. The following conclusions can be reached from this study.
Compared to the original YOLOv7, this algorithm structure has achieved a certain degree of improvement in all four metrics. Therefore, it is feasible to achieve better detection results through improvements. The success of the experiment has laid a solid foundation for further optimization of detection models in planetary analysis.
At the same time, we found that although the new algorithm performs better than the original YOLOV7 structure in detecting small impact crater and crater with missing edges, it is still easily misjudged. This is also a challenge we need to address and improve in the future. Additionally, we will also explore the effectiveness of the new structure in detecting Mars crater and investigate the transferability of the structure.
Acknowledgements
This work was supported by the Science and Technology Development Fund of Macau (grant number 0020/2024/RIA1).
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, author-ship, and/or publication of this article.
Data Sharing Agreement
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.
References
[1] Guardabasso, P.; Paternostro, S.; Bedialauneta, P.; Fonteyne, R. Lunar landing necessary building blocks and good practices for a sustainable development of human lunar activities. Acta Astronautica 2023, 202, 782–790.
[2] Barker, D.C. Lunar and off Earth resource drivers, estimations and the development conundrum. Advances in Space Research 2020, 66, 359–377.
[3]ood, C.A.; Anderson, L. New morphometric data for fresh lunar craters. In Proceedings of the In: Lunar and Planetary Science Conference, 9th, Houston, Tex., March 13-17, 1978, Proceedings.(A79-39253 16-91) New York, Pergamon Press, Inc., 1978, p. 3669-3689., 1978, Vol. 9, pp. 3669–3689.
[4] Robbins, S.J. A new global database of lunar impact craters> 1–2 km: 1. Crater locations and sizes, comparisons with published databases, and global analysis. Journal of Geophysical Research: Planets 2019, 124, 871–892.
[5]. Lu, Y.; Hu, Y.; Xiao, J.; Liu, L.; Zhang, L.; Wang, Y. Three-Dimensional Model of the Moon with Semantic Information of Craters Based on Chang’e Data. Sensors 2021, 21, 959.
[6] Zhang, K.; Liu, J.; Zhang, L.; Gu, Y.; Yue, Z.; Zhang, S.; Zhang, J.; Qin, H.; Liu, J. Quantitative Research on the Morphological Characteristics of Lunar Impact Craters of Different Stratigraphic Ages since the Imbrian Period. Remote Sensing 2024, 16, 1540.
[7] Pedrosa, M.M.; de Azevedo, S.C.; da Silva, E.A.; Dias, M.A. Improved automatic impact crater detection on Mars based on morphological image processing and template matching. Geomatics, Natural Hazards and Risk 2017, 8, 1306–1319.
[8] Saheba, S.M.; Upadhyaya, T.K.; Sharma, R.K. Lunar surface crater topology generation using adaptive edge detection algorithm. IET Image Processing 2016, 10, 657–661.
[9] Wang, Y.W.; Qin, C.Z.; Cheng, W.M.; Zhu, A.X.; Wang, Y.J.; Zhu, L.J. Automatic Crater detection by training random forest classifiers with legacy Crater map and spatial structural information derived from digital terrain analysis. Annals of the American Association of Geographers 2022, 112, 1328–1349.
[10] Ren, S.; He, K.; Girshick, R.; Sun, J. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 2015, 28.
[11]Zou, C.; Lai, J.; Liu, Y.; Cui, F.; Xu, Y.; Qiao, L. Small lunar crater identification and ageestimation in Chang’e-5 landing area based on improved Faster R-CNN. Icarus 2024, 410, 115909.
[12] Jiang, S.; Lian, Z.; Yung, K.L.; Ip, W.H.; Gao, M. Automated detection of multitype landforms on mars using a light-weight deep learning-based detector. IEEE Transactions on Aerospace and Electronic Systems 2022, 58, 5015–5029.
[13] Qiu, Y.; Liang, Y.; Chen, X.; Zhang, Z.; Xie, S.; Zhou, G. Lightweight tensorial convolutional neural network for lunar impact crater detection. Advances in Space Research 2024, 74, 518–527.
[14] Wang, X., Zhang, Z., & Shang, X. Research on Improved YOLO11 for Detecting Small Targets in Sonar Images Based on Data Enhancement. Applied Sciences, 15(12), 6919.
[15] Xiong, Z., Wang, M., Kan, R., & Zhang, J. YOLO-SAR: An Enhanced Multi-Scale Ship Detection Method in Low-Light Environments. Applied Sciences, 15(13), 7288.
[16] Wang, C.Y.; Bochkovskiy, A.; Liao, H.Y.M. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 7464–7475.
[17] Dai, J.; Qi, H.; Xiong, Y.; Li, Y.; Zhang, G.; Hu, H.; Wei, Y. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, 2017, pp. 764–773.
[18] Zheng, Z.;Wang, P.; Ren, D.; Liu,W.; Ye, R.; Hu, Q.; Zuo,W. Enhancing geometric factors in model learning and inference for object detection and instance segmentation. IEEE transactions on cybernetics 2021, 52, 8574–8586.