Complexity Problems Handled by Advanced Computer Simulation Technology in Smart Cities 2020View this Special Issue
Multiobject Detection Algorithm Based on Adaptive Default Box Mechanism
Multiobject detection tasks in complex scenes have become an important research topic, which is the basis of other computer vision tasks. Considering the defects of the traditional single shot multibox detector (SSD) algorithm, such as poor small object detection effect, reliance on manual setting for default box generation, and insufficient semantic information of the low detection layer, the detection effect in complex scenes was not ideal. Aiming at the shortcomings of the SSD algorithm, an improved algorithm based on the adaptive default box mechanism (ADB) is proposed. The algorithm introduces the adaptive default box mechanism, which can improve the imbalance of positive and negative samples and avoid manually set default box super parameters. Experimental results show that, compared with the traditional SSD algorithm, the improved algorithm has a better detection effect and higher accuracy in complex scenes.
With the continuous improvement of deep learning related theories, computer vision technologies [1–3] have achieved great success. As the basis of computer vision tasks, object detection [4–6] has been applied in many fields such as intelligent security , automatic driving , and intelligent medical treatment ; even some industrial applications are based on object detection algorithms [10–13]. In the past few years, in order to improve the real-time performance and accuracy of object detection in complex scenes, many scholars have conducted a lot of research on this, and the object detection algorithms based on deep learning have achieved remarkable achievements.
In 2014, the Region with CNN features (RCNN) algorithm  was published in Computer Vision and Pattern Recognition (CVPR) by Ross Girshick et al. The advent of this algorithm marked a new era in object detection technology. After that, Spatial Pyramid Pooling in Deep Convolutional Networks (SPPNet) algorithm  makes up for the shortcomings of the RCNN algorithm in repetitive convolution calculation and fixed output scale, but it still has the defects of tedious training steps and slow process. In order to improve the real-time performance of the RCNN detection algorithm, Ross Girshick proposed the Fast RCNN object detection algorithm  in 2015. The processing mode of shared convolution makes the calculation amount of this algorithm drop sharply. In addition, the method of the region of interest (RoI) pooling is introduced to enable the network to process input images of any size. However, in this method, the problem of time loss caused by the selective search method  has not been solved. As a result, the Faster RCNN  algorithm was published in Neural Information Processing Systems (NIPS) in 2015. The highlight of this algorithm is to propose region proposal network (RPN) network structure, which combines region generation with convolution neural network based on the default box mechanism. It further improves the real-time performance of the Fast RCNN algorithm and becomes the most representative algorithm in the two-stage detection algorithm. Based on the Faster RCNN algorithm, Mask RCNN , fully convolutional network Region-based Fully Convolutional Networks (R-FCN) , Cascade RCNN , and other improved algorithms were proposed. Compared with the two-stage structure of Faster RCNN and other algorithms, the You Only Look Once (YOLO) series of algorithms [22–25] and the SSD algorithm  adopt the single-stage structure to directly predict the location and category of the object. The real-time performance of the algorithm is greatly improved, but the accuracy of object detection in complex scenes is obviously insufficient. In order to improve the accuracy of the single-stage detection algorithm, Fu et al.  proposed a feature fusion method for multiscale prediction based on the SSD algorithm and used deconvolution operation to enhance the semantic information of shallow features. Jeong et al.  tried pooling fusion, deconvolution fusion, rainbow fusion, and other schemes and finally designed the Rainbow Single Shot Multibox Detector (RSSD) network model, which improved the object detection effect in complex scenes to some extent. Tsung et al.  considered that the reason for the lack of accuracy of the single-stage detection model was the imbalance of positive and negative samples and proposed a new classification loss function “Focal Loss” to improve the problem. Based on Focal Loss, the team of Tsung built a RetinaNet network, which effectively balanced the proportion of positive and negative samples to avoid the imbalance of samples.
This paper is based on the SSD algorithm, and we improve it from the following two aspects in view of the shortcomings of the SSD algorithm. On the one hand, in order to enhance the characterization capability of the low feature layers and the detection effect of small objects, the improved algorithm introduces feature layers fusion (FLF) and multireceptive field fusion (MRFF) mechanisms. On the other hand, through the adaptive default box mechanism, the steps of manually setting default box hyperparameters are avoided, the generation of negative sample box is reduced, and the problem of positive and negative sample imbalance is improved. Under the premise of real-time detection, the improved algorithm greatly improves the accuracy of small object detection in complex scenes.
2. Related Work
2.1. SSD Algorithm
The traditional SSD algorithm takes Visual Geometry Group Network (VGGNet)  as the backbone network and adds several additional convolution layers to participate in the detection of related objects. Firstly, sufficient data enhancements have been made through optical changes, geometric transformations, etc., which greatly enriched the relevant data sets. Secondly, the SSD algorithm expands 4 convolutional layers and performs object detection based on convolutional layers of different depths. Therefore, the output feature maps have different scales and receptive fields, and objects of different sizes can be detected. Thirdly, the SSD algorithm sets multiple default boxes of fixed size and ratio on the six feature maps. The algorithm sets a series of smaller default boxes on the shallow feature maps to detect small objects and sets several larger default boxes on the deep feature maps to detect large objects. Finally, the network model uses 3 × 3 convolution kernels to extract features on the relevant feature maps to complete objects classification and bounding boxes regression.
The SSD algorithm completes object detection through the single-stage network and has a better effect compared with the algorithm in the same period. Correspondingly, the SSD algorithm also has some shortcomings. On the one hand, due to the lack of semantic information in the shallow feature maps, the classification and regression effect of small objects is poor and the detection accuracy is insufficient. On the other hand, the default box parameters of each feature layer depend on the manual setting, so the generalization of the SSD is poor in different detection tasks.
2.2. Design Criteria and Defects of Default Box in Traditional Detector
A series of default boxes are generated by using the sliding window method in the relevant feature maps of the models, which is the mainstream method adopted by various object detection models at present. Firstly, the model defines several default boxes with specific scales and aspect ratios. Secondly, a large number of default boxes for object detection tasks are generated by sliding in the relevant output feature maps with a certain step size. The traditional default box generation method has the following disadvantages:(1)The traditional object detection models need to define a series of aspect ratios and scales for the default box. The selection of the default box aspect ratios and scales of the model will directly affect the detection effect of the models. In addition, for different data sets and detection methods, the parameters of the default box need to be adjusted according to the situation. If the selected default box parameters are not appropriate, the recall rate of the model will be too low, and the detection effect of object detection model will be poor(2)In the output feature maps of the relevant models, a large number of default boxes are distributed in the background area of the input image, which cannot play a good role in the detection of relevant objects(3)For the objects with a large difference in size and aspect ratio, a series of predefined default boxes may not be able to meet the detection requirements of the model(4)A large number of default boxes will directly lead to the degradation in the precision rate and real-time performance of the detection model
3. Improved Algorithm Design
3.1. Fusion Mechanism in the Improved Algorithm
In order to fuse the low-level output feature maps of different scales, on the one hand, the feature maps of Conv4_3, FC7, and Conv6_2 were dimensionally reduced to 256. The feature maps of FC7 and Conv6_2 were adjusted to 38 × 38 by bilinear interpolation method and concat operation is carried out for the relevant feature maps after processing. On the other hand, the dimension of the FC7 feature map is reduced to 512, the dimension of the Conv6_2 output feature map remains unchanged, and the dimension of Conv7_2 is transformed to 512. The bilinear interpolation method is used to adjust the output feature maps of Conv6_2 and Conv7_2 to 19 × 19. Similarly, the concat operation is performed based on the processed feature maps.
The size of the human receptive field will change with the eccentricity of retinal imaging, and the size of the receptive field is proportional to the eccentricity. In order to further improve the detection efficiency of the traditional SSD object detection model, the multireceptive field fusion mechanism was added to the improved model by referring to the human visual perception mechanism.
Based on convolution kernels of different sizes and dilated convolution of different scales , the relevant mechanism fuses multiple scales of receptive fields to make the improved model have stronger feature expression ability. The fusion mechanism of multiple receptive fields in the improved model is shown in Figure 1.
The multireceptive field fusion mechanism consists of three branches. Firstly, convolution kernels of 1 × 1, 3 × 3, and 5 × 5 are used to simulate receptive fields of different sizes. Secondly, dilated convolution with rates of 1, 3, and 5 is used to simulate different degrees of eccentricity. In addition, for the 3 × 3 and 5 × 5 branches, the dimension of the feature map is reduced by using 1 × 1 convolution kernel, so as to reduce the number of parameters. The feature map after dimension reduction is then sent into the convolution kernels of 3 × 3 and 5 × 5. Finally, the fusion of 3 branches is completed by channel concat, and the number of feature channels is reduced by 1 × 1 convolution kernel.
3.2. Adaptive Default Box Mechanism
The parameters’ setting of the default box is the key part of the SSD object detection method. Similar to most mainstream object detection methods, the setting and generation of default boxes in the SSD algorithm also rely on the artificial unified setting. A series of preset default boxes are applied to the output features of relevant detection layers in the SSD algorithm. Since there are a large number of default boxes in the background area of the input image, and the aspect ratio of predefined default boxes may not be applicable to the objects to be detected in the relevant image, therefore, the detection efficiency of the model is greatly reduced by using this scheme.
The distribution of the objects in the input image is usually uneven, and the generation of default boxes is usually related to the content of the input image, the location, and the shape of the objects to be detected. Accordingly, the improved SSD object detection algorithm no longer uses the traditional default box generation strategy. The semantic information obtained by the algorithm is used to guide the generation of a series of appropriate size default boxes. The default box of an object is represented as (x, y, , h), where (x, y) represents the central coordinate position, and and h represent the width and height, respectively. Assuming that A is an object to be detected on the input image G, the distribution of the corresponding default box can be represented by
According to equation (1), we can obtain two aspects of information. On the one hand, the object A to be detected may only appear in a partial area of the input image G. On the other hand, the distribution and scale of the corresponding default box are closely related to the location of object A. Therefore, the adaptive default box mechanism of the improved SSD model is shown in Figure 2.
The adaptive default box generation mechanism includes two parts: position prediction and shape prediction. Assuming that the input image is G, on the one hand, the position feature map is generated through the position prediction branch of the mechanism. The probability and position distribution of the objects to be detected in the input image can be obtained through the position feature map. On the other hand, according to the position prediction and shape prediction branches, the sizes and aspect ratios of the default boxes are predicted to generate the default boxes with different sizes and aspect ratios. Therefore, the default boxes in the improved SSD model are variable, and different contents can be obtained according to the features in different positions of the output feature maps. Considering that the shape of the default box is not fixed, by introducing the feature adaptive module, we carry out the feature adaptive adjustment for the improved model.
3.3. Default Box Position and Shape Prediction
In the process of position prediction, the improved SSD detection model first generates a series of location feature maps.
We assume that is the coordinate of a point in the position feature map, and its probability value P corresponds to the coordinate Q in the input image, which can be expressed by
in which represents the output feature map of a certain detection layer and s represents the step size of the output feature map. The 1 × 1 convolution kernel is used to process the output feature map of the relevant detection layer, and the score map of the objects to be detected in the input image is obtained. The position prediction map of is further generated by the Sigmoid function. A certain probability threshold is set to identify the possible position of the object to be detected.
Based on the position prediction of the default box, the default bounding boxes of the objects to be detected are predicted by the shape prediction branch. According to the output feature map of the relevant detection layer, the shape prediction branch of the default box will predict the best default box shape at each location in the feature map. That is, by predicting the width and height of the default box, the maximum IoU value can be generated as far as possible with the nearest ground truth bounding box. Due to the fact that the range generated when directly predicting the width and height of the bounding box is wide, and the prediction result is unstable, so it can be converted bytwhere s represents the step size and represents the relative parameters controlling the default box size. Through equation (3), the output space can be mapped from [0, 1000] to [−1, 1], so that the improved SSD object detection model can detect relevant objects more stably. The shape prediction branch uses the convolutional kernel of 1 × 1 to predict the dw and dh values of the default box and completes the pixel-level transformation of the relevant feature map through equation (3).
Compared with SSD, YOLO, RSSD, DSSD, and other object detection models, on the one hand, each position in the traditional models corresponds to a set of preset default bounding boxes. Each position in the feature maps of the improved model corresponds to only one prediction default box. The number of default boxes is greatly reduced, and the generated default boxes are more closely related to the objects to be detected. On the other hand, in the default box prediction scheme of the improved model, the aspect ratio of the default box does not need to be set manually. So, it also has a better detection effect for the abnormal size objects existing in the input image.
3.4. Feature Adaptive Module
In most object detection networks such as SSD, RSSD, and DSSD, the sizes and aspect ratios of the default box are consistent at each position of the feature map. Therefore, the general convolution can be used to extract features in the output feature maps of the detection layers. Furthermore, the relevant features of each default bounding box are expressed. Compared with the existing SSD, RSSD, DSSD, and other object detectors, the default boxes with different shapes are automatically generated in the improved model. The output feature maps of the detection layers cannot predict the shape of its default boxes, but it is necessary to predict the categories and position offsets of these default boxes in the subsequent stages. That is to say, there is a mismatch between the default box and the features of the default box in the improved SSD model. In order to solve the above problem, the improved model introduces the relevant feature adaptive module and adjusts the relevant output feature maps according to equation (4) based on the default box shape of each position:where represents the feature at the ith position in the output feature map and represents the width and height of the default box corresponding to the ith position. After the prediction of the default box, in order to realize the relevant position transformation and adapt to the shape of the default bounding box, a 3 × 3 deformable convolution is applied to the output feature map to realize NT . Different from the ordinary deformable convolution, the bias value in the feature adaptive module comes from the predicted default boundary box; that is, 1 × 1 convolution kernel is used to act on the predicted default bounding box. From the perspective of specific functions, the feature adaptive module of the improved SSD model is similar to the RoI Pooling layer in the Faster RCNN algorithm. The structure of the improved SSD model is shown in Figure 3.
3.5. Loss Function Setting
Different from the traditional object detection models, the loss in the improved model includes not only the general classification loss and regression loss but also the position loss and shape loss during the default bounding box prediction. The final loss function can be expressed by equation (5), and the position loss and shape loss are balanced by the parameters and :where the classification loss adopts Cross Entropy (CE) loss  and the regression loss adopts smooth L1 loss. and can be expressed by where represents the probability that sample i is predicted to be of a certain class. indicates that the ith sample belongs to a label of a certain category, and its value is 0 or 1. l and g, respectively, represent the deviation between the prediction box and the ground true box with the default box. When the default boundary box is generated, since the number of positive samples is smaller than that of negative samples, the focal loss is adopted to solve the problem of unbalanced positive and negative samples in position prediction. It can effectively reduce the loss of positive samples and the weight of negative samples in the training process. The loss can be expressed in equation (8). The value of the balance factor is set to 0.25 and the value of the regulation coefficient is set to 2:
When calculating the shape loss of the model, IoUmax is taken as a measure of the relevant loss. Based on the position feature map of each detection layer, several groups of different aspect ratios are sampled at each positive sample point position to complete the matching of IoU and determine the optimization object. Correlation matching can be expressed by equation (9). The shape loss of the improved model is shown in equation (10), where , hp, , and , respectively, represent the shapes of the prediction bounding box and the real bounding box:
4. Experimental Exploration
4.1. Experimental Data Sets
In view of the distribution of the objects in complex scenes, the ideal object detection model should have good generalization performance, which can not only effectively detect all kinds of different sizes of objects but also stably detect dense scenes. Based on Crowd Human , Pascal VOC 2012, and MS COCO data sets , relevant experiments verify the detection efficiency and robustness of the model on different data sets and comprehensively evaluate the improved model.
The Crowd Human data set was released by MEGVII in 2018. The training set and validation set contain about 20,000 images. The data set contains more than 470,000 instances, and there are about 23 human instances in each image on average. At the same time, complex occlusion phenomenon exists in the images.
PASCAL VOC is a classic public data set, which mainly includes VOC 2007 and VOC 2012. This data set provides a complete set of standards for image classification and object detection. The VOC 2007 data set is extended to the VOC 2007 data set, which contains 20 kinds of objects and 11530 images. The related image annotations are complete and of high quality.
MS COCO data set is funded and annotated by Microsoft. It involves multiple computer vision tasks such as object detection, object segmentation, and semantic understanding. It contains about 300,000 data images, more than 2 million instances, and 91 kinds of objects. Compared with other public data sets, the COCO data set has more small objects, more complex object types, and detection scenarios. It can comprehensively evaluate model performance.
4.2. Data Preprocessing and Model Evaluation Indexes
In order to fully train the improved model, enhance the generalization of the model, and improve the detection effect of small objects and occlusion objects, the corresponding data preprocessing strategy is formulated. Generally, the object whose number of pixels is less than 1024 in the segmentation mask of the image object region is defined as a small object. Objects with more than 1024 pixels and less than 9216 pixels in the segmentation mask of the image object region are defined as medium-size objects. It mainly includes two aspects: optical transformation and geometric transformation. Optical transformation mainly includes the adjustment of brightness, contrast, hue, saturation, and channel. The geometric transformation utilizes operations such as random cropping, random expansion, and scaling to achieve image size changes.
The performance of the improved model is measured by average precision (AP), average recall (AR), and frame per second (FPS). As the common evaluation indexes, the AP value reflects the precision and recall rate of the test results. The larger the value is, the better the detection precision of the model will be, and the AR value reflects the recall rate and positioning accuracy of the model. In addition to the detection precision, the FPS value is used to measure the detection speed of the improved algorithm, that is, the number of images the model can process per second.
4.3. Model Parameters’ Setting and Training
The relevant models are trained and tested on Crowd Human, PASCAL VOC 2012, and MS COCO data sets, respectively, to verify the generalization performance of the improved model on different data sets. In the multitask loss function, and are set to balance the position loss and shape loss of the default box. The training of the model is based on the stochastic gradient descent algorithm and the “warm-up” strategy is adopted. During the initial five epochs, the learning rate of the model is increased from 10−4 to 410−3. After the “warm-up” phase, the learning rate is changed to 10−4 again, and the learning rate is set as 10−5 and 10−6, respectively, at the 8th epoch and the 11th epoch. The momentum value during the training process is 0.9, and the weight attenuation value is 0.005. All experiments are conducted based on GTX 1080 Ti and Titan X GPU environments.
4.4. Influence of Relevant Mechanisms on the Detection Effect
The improved SSD detection model is based on multiple feature maps for object detection. The deep feature maps with a large receptor field are responsible for the detection of large-scale objects, while the low-level feature maps with a small receptor field are responsible for the detection of small objects. By introducing the corresponding fusion mechanisms, the semantic information of low feature layers can be enriched. Accordingly, Conv 4_3, Conv7_fc, Conv F1, and Conv F2 are used for the detection of small objects, while the rest of the detection layers are used for the detection of larger objects. In addition, the ADB mechanism is added to the improved SSD model to improve the positioning precision of the model, avoid manually setting default box hyperparameters, and improve the imbalance of positive and negative samples. Based on different data sets, the experiment explored the influence of relevant mechanisms on the detection effect, and the experimental results are shown in Table 1.
Based on different data sets, Table 1 explores the influence of relevant mechanisms on the detection results of Conv 4_3, Conv7_fc, Conv F1, and Conv F2 layers. For Conv4_3 layer, ADB mechanism was added to the improved algorithm, and the AP values of Crowd Human, PASCAL VOC 2012, and MS COCO data sets reached 92.1%, 72.6%, and 45.3%, respectively. AR value can be up to 81.6%, 61.4%, and 36.1%; compared with the traditional SSD algorithm, the average precision value and the average recall rate are greatly improved. In addition, it can be seen that the detection effect of Conv7_fc has also been significantly improved.
In order to enhance the detection effect of small objects in dense scenes, additional small object detection layers Conv F1 and Conv F2 are added in the improved algorithm. The relevant detection layers use FLF, MRFF, and ADB generation mechanism to strengthen the semantic information of the low detection layers. In the case of applying FLF and MR, the average detection precision of the Conv F1 detection layer on the three data sets can reach 87.5%, 68.3%, and 41.2%, respectively. Compared with the detection effect of the Conv4_3 layer in the traditional SSD algorithm, the algorithm precision is improved. After the introduction of the ADB mechanism, the average detection precision and average recall rate of the algorithm are greatly improved. The experiment shows that the improved network has stronger characterization ability, better detection effect, and higher object positioning precision. Figure 4 shows the influence of relevant mechanisms on the detection effect. With the introduction of ADB, FLF, and MR, the low detection layers of the improved algorithm can extract richer feature information and detect more small objects compared with the original algorithm.
4.5. Comparison of Relevant Models
Based on the PASCAL VOC2012 test set, we compared the detection effects of Faster RCNN, YOLO V2, SSD, DSSD, RSSD, and our SSD algorithms. The training of the algorithm involved VOC 2012 and MS COCO training sets. The basic network included VGGNet, ResNet-101 , and Darknet-19. Taking FPS, mAP, and mAR  as evaluation criteria, the experimental comparison results of the six models are shown in Table 2.
By analyzing the experimental data in Table 2, our SSD300 has improved its average precision and average recall rate compared with Faster RCNN, YOLOv2, SSD300, DSSD321, and RSSD300 algorithms. The detection precision of our SSD300−S0 model can reach 73.2% without pretraining, which is 3.6% higher than that of SSD300−S0. When the model training is combined with the MS COCO data set, the detection accuracy of our SSD300+coco reaches 83.4%, which is 2.2% higher than SSD300+coco. In addition, the average recall rate of our SSD300+coco is 74.1%, which is about 2.5% higher than SSD300+coco. This verifies the effectiveness of ADB and other relevant mechanisms, improves the imbalance of positive and negative samples in traditional SSD algorithms, and improves the detection effect of objects in dense scenes.
Due to the introduction of ADB and other modules, some detection time is lost. Compared with the original SSD300, the real-time detection performance of the improved algorithm is reduced, but it is enough to meet the requirements of real-time detection. Compared with RSSD300, DSSD321, and other improved algorithms, our SSD300 not only improves the average detection precision and average recall rate but also has obvious advantages in real-time detection.
In order to further verify the generalization of the improved algorithm on different data sets, a series of AP values and AR values of the relevant models were fully explored based on the MS COCO data set. The relevant experimental results are shown in Tables 3 and 4.
According to the experimental data in Tables 3 and 4, compared with Faster RCNN, YOLO V2, SSD, DSSD, and RSSD algorithms, our SSD still has good detection performance on MS COCO data set. In the detection of small objects, APS and ARS of our SSD512 can reach 14.3% and 23.6%, respectively. Compared with the original SSD algorithm, the average detection precision and average recall rate of small objects have been improved by about 3.4% and 7.1%, respectively. In addition, the other evaluation indicators also have different degrees of improvement. The improved algorithm achieves ideal detection results on both MS COCO and PASCAL VOC data sets. On the one hand, the improved SSD algorithm has good generalization. On the other hand, it also directly shows the effectiveness of the algorithm improvement.
In view of the defects of the traditional SSD detection algorithm, such as the poor detection effect of small objects and the default box generation depending on manual settings, this paper proposes an improved multiobject detection algorithm, which effectively improves the object detection effect in complex scenes. The improved algorithm mainly involves the following contributions: on the one hand, the introduction of feature fusion and multireceptive field fusion mechanism enhances the characterization ability of the low feature layers and improves the detection effect of small objects. On the other hand, through the adaptive default box mechanism, the steps of setting default box hyperparameters are avoided, the generation of negative sample box is reduced, and the imbalance of positive and negative samples is improved. Under the requirement of real-time detection, the improved algorithm greatly improves the average precision and recall rate of object detection in complex scenes.
The image data supporting this algorithm research is from previously reported studies and datasets. These prior studies and datasets are cited at relevant places within the text as references [22–24, 29, 30].
Conflicts of Interest
The authors declare that they have no conflicts of interest in this work.
All authors contributed equally to this manuscript.
This work was supported in part by the Soft Science Foundation of Shanxi Province under Grant 2011041037-02 and in part by the Scholarship Council of Shanxi Province under Grant 2011-8.
S. Hussein, P. Kandel, C. W. Bolan, M. B. Wallace, and U. Bagci, “Lung and pancreatic tumor characterization in the deep learning era: novel supervised and unsupervised learning approaches,” IEEE Transactions on Medical Imaging, vol. 38, no. 8, pp. 1777–1787, 2019.View at: Publisher Site | Google Scholar
R. Girshick, J. Donahuej, T. Darrell, and T. Malik, “Rich feature hierarchies for accurate object detection and semantic segmentation,” in Proceedings of the Conference on Computer Vision and Pattern Recognition, pp. 580–587, IEEE, Columbus, OH, USA, June 2014.View at: Google Scholar
R. Girshick, “Fast R-CNN,” in Proceedings of the International Conference on Computer Vision, pp. 1440–1448, IEEE, Santiago, Chile, December 2015.View at: Google Scholar
S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: towards real-time object detection with region proposal networks,” in Proceedings of the Advances in Neural Information Processing Systems, pp. 91–99, Curran Associates Inc, Montreal, Canada, December 2015.View at: Google Scholar
K. He, G. Gkioxari, P. Dollár, and R. Girshick, “Mask R-CNN,” in Proceedings of the International Conference on Computer Vision, pp. 2961–2969, IEEE, Venice, Italy, October 2017.View at: Google Scholar
J. Dai and Y. Li, “R-FCN: object detection via region-based fully convolutional networks,” in Proceedings of the Advances in Neural Information Processing Systems, vol. 29, pp. 379–387, NIPS, Barcelona, Spain, December 2016.View at: Google Scholar
Z. Cai and N. Vasconcelos, “Cascade R-CNN: delving into high quality object detection,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6154–6162, IEEE, Salt Lake City, UT, USA, June 2018.View at: Google Scholar
J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: unified, real-time object detection,” in Proceedings of the Computer Vision and Pattern Recognition, pp. 779–788, IEEE, LasVegas, NV, USA, June 2016.View at: Google Scholar
J. Redmon and A. Farhadi, “YOLO9000: better, faster, stronger,” in Proceedings of the Computer Vision and Pattern Recognition, pp. 7263–7271, IEEE, Honolulu, HI, USA, July 2017.View at: Google Scholar
W. Liu, D. Anguelov, D. Erhan et al., “SSD: single shot multibox detector,” in Proceedings of the European Conference on Computer Vision, pp. 21–37, Springer, Cham, Switzerland, October 2016.View at: Google Scholar
J. Jeong, H. Park, and N. Kwak, “Enhancement of SSD by concatenating feature maps for object detection,” in Proceedings of the British Machine Vision Conference, BMVA Press, London, UK, May 2017.View at: Google Scholar
T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal loss for dense object detection,” in Proceedings of the International Conference on Computer Vision, pp. 2980–2988, IEEE, Venice, Italy, October 2017.View at: Google Scholar
K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” in Proceedings of the International Conference on Learning Representations 2015, San Diego, CA, USA, May 2015.View at: Google Scholar
G. E. Nasr, E. A. Badr, and C. Joun, “Cross entropy error function in neural networks: forecasting gasoline demand,” in Proceedings of the Fifteenth International Florida Artificial Intelligence Research Society Conference, AAAI Press, Pensacola Beach, FL, USA, May 2002.View at: Google Scholar
T. Y. Lin, M. Maire, S. Belongie et al., “Microsoft COCO: common objects in context,” in Proceedings of the European Conference on Computer Vision, pp. 740–755, Springer, Zurich, Switzerland, September 2014.View at: Google Scholar
K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the Computer Vision and Pattern Recognition, pp. 770–778, IEEE, LasVegas, NV, USA, June 2016.View at: Google Scholar