Ground-truth Boxes Research Articles

The Lightweight Anchor Dynamic Assignment Algorithm for Object Detection.

Smart security based on object detection is one of the important applications of edge computing in IoT. Anchors in object detection refer to points on the feature map, which can be used to generate anchor boxes and serve as training samples. Current object detection models do not consider the aspect ratio of the ground-truth boxes in anchor assignment and are not well-adapted to objects with very different shapes. Therefore, this paper proposes the Lightweight Anchor Dynamic Assignment algorithm (LADA) for object detection. LADA does not change the structure of the original detection model; first, it selects an equal proportional center region based on the aspect ratio of the ground-truth box, then calculates the combined loss of anchors, and finally divides the positive and negative samples more efficiently by dynamic loss threshold without additional models. The algorithm solves the problems of poor adaptability and difficulty in the selection of the best positive samples based on IoU assignment, and the sample assignment for eccentric objects and objects with different aspect ratios was more reasonable. Compared with existing sample assignment algorithms, the LADA algorithm outperforms the MS COCO dataset by 1.66% over the AP of the baseline FCOS, and 0.76% and 0.24% over the AP of the ATSS algorithm and the PAA algorithm, respectively, with the same model structure, which demonstrates the effectiveness of the LADA algorithm.

Improving YOLOv7-Tiny for Infrared and Visible Light Image Object Detection on Drones

To address the phenomenon of many small and hard-to-detect objects in drone images, this study proposes an improved algorithm based on the YOLOv7-tiny model. The proposed algorithm assigns anchor boxes according to the aspect ratio of ground truth boxes to provide prior information on object shape for the network and uses a hard sample mining loss function (HSM Loss) to guide the network to enhance learning from hard samples. This study finds that the aspect ratio difference of vehicle objects under drone perspective is more obvious than the scale difference, so the anchor boxes assigned by aspect ratio can provide more effective prior information for the network than those assigned by size. This study evaluates the algorithm on a drone image dataset (DroneVehicle) and compares it with other state-of-the-art algorithms. The experimental results show that the proposed algorithm achieves superior average precision values on both infrared and visible light images, while maintaining a light weight.

ReBiDet: An Enhanced Ship Detection Model Utilizing ReDet and Bi-Directional Feature Fusion

To enhance ship detection accuracy in the presence of complex scenes and significant variations in object scales, this study introduces three enhancements to ReDet, resulting in a more powerful ship detection model called rotation-equivariant bidirectional feature fusion detector (ReBiDet). Firstly, the feature pyramid network (FPN) structure in ReDet is substituted with a rotation-equivariant bidirectional feature fusion feature pyramid network (ReBiFPN) to effectively capture and enrich multiscale feature information. Secondly, K-means clustering is utilized to group the aspect ratios of ground truth boxes in the dataset and adjust the anchor size settings accordingly. Lastly, the difficult positive reinforcement learning (DPRL) sampler is employed instead of the random sampler to address the scale imbalance issue between objects and backgrounds in the dataset, enabling the model to prioritize challenging positive examples. Through numerous experiments conducted on the HRSC2016 and DOTA remote sensing image datasets, the effectiveness of the proposed improvements in handling complex environments and small object detection tasks is validated. The ReBiDet model demonstrates state-of-the-art performance in remote sensing object detection tasks. Compared to the ReDet model and other advanced models, our ReBiDet achieves mAP improvements of 3.20, 0.42, and 1.16 on HRSC2016, DOTA-v1.0, and DOTA-v1.5, respectively, with only a slight increase of 0.82 million computational parameters.

Lightweight detection networks for tea bud on complex agricultural environment via improved YOLO v4

Rapid and accurate detection of tender tea buds in the natural tea garden environment is the basis for intelligent tea picking. However, complex models result in high hardware computing power requirements and limit the deployment of tea bud recognition models in tea-picking robots. Therefore, this paper investigates a high-precision and lightweight target detection model based on improved you-only-look-once version 4(YOLOv4). The lightweight network GhostNet is used to replace the backbone of YOLOv4, and the depthwise separable convolution is designed to substitute the standard convolution, significantly reducing the computational load and computational complexity of the model. Additionally, the convolution block attention module (CBAM) is embedded into the path aggregation network (PANet), which enhances the model’s feature extraction capability. To solve the problem of detection and distinction caused by overlapping of one-bud-one-leaf and one-bud-two-leaf, this paper proposes the CIoU loss function in YOLOv4 to the SIoU loss function. The SIoU loss function considers the vector angles of the ground truth box and the prediction box and redefines the penalty indicator to improve the training speed and detection accuracy of the model. The experimental results show that the detection accuracy of the proposed approach is 85.15% for the one-bud-one-leaf and one-bud-two-leaf. The giga floating point operations per second (GFlops) and parameters are 6.594 G and 11.353 M. Relative to the original YOLOv4, the proposed algorithm’s mean accuracy is improved by 1.08%, 89.11% reduces the computational complexity, and 82.36% reduces the number of parameters. The Tea-YOLO algorithm demonstrates significantly better actual detection performance in different angles and natural environments compared to the YOLOv4 algorithm. The algorithm proposed in this paper can detect one-bud-one-leaf and one-bud-two-leaf quickly and accurately, which reduces the cost and difficulty of deploying the vision module of the tea-picking robot.

Location First Non-Maximum Suppression for Uncovered Muck Truck Detection

Uncovered muck truck detection aims to detect the muck truck and distinguish whether it is covered or not by dust-proof net to trace the source of pollution. Unlike traditional detection problem, recalling all uncovered trucks is more important than accurate locating for pollution traceability. When two objects are very close in an image, the occluded object may not be recalled because the non-maximum suppression (NMS) algorithm can remove the overlapped proposal. To address this issue, we propose a Location First NMS method to match the ground truth boxes and predicted boxes by position rather than class identifier (ID) in the training stage. Firstly, a box matching method is introduced to re-assign the predicted box ID using the closest ground truth one, which can avoid object missing when the IoU of two proposals is greater than the threshold. Secondly, we design a loss function to adapt the proposed algorithm. Thirdly, a uncovered muck truck detection system is designed using the method in a real scene. Experiment results show the effectiveness of the proposed method.

PointPainting: 3D Object Detection Aided by Semantic Image Information.

A multi-modal 3D object-detection method, based on data from cameras and LiDAR, has become a subject of research interest. PointPainting proposes a method for improving point-cloud-based 3D object detectors using semantic information from RGB images. However, this method still needs to improve on the following two complications: first, there are faulty parts in the image semantic segmentation results, leading to false detections. Second, the commonly used anchor assigner only considers the intersection over union (IoU) between the anchors and ground truth boxes, meaning that some anchors contain few target LiDAR points assigned as positive anchors. In this paper, three improvements are suggested to address these complications. Specifically, a novel weighting strategy is proposed for each anchor in the classification loss. This enables the detector to pay more attention to anchors containing inaccurate semantic information. Then, SegIoU, which incorporates semantic information, instead of IoU, is proposed for the anchor assignment. SegIoU measures the similarity of the semantic information between each anchor and ground truth box, avoiding the defective anchor assignments mentioned above. In addition, a dual-attention module is introduced to enhance the voxelized point cloud. The experiments demonstrate that the proposed modules obtained significant improvements in various methods, consisting of single-stage PointPillars, two-stage SECOND-IoU, anchor-base SECOND, and an anchor-free CenterPoint on the KITTI dataset.

Fast object detector with center localization confidence based on FCOS for environment perception in urban traffic scene

Environment perception on urban roads is one of the most important tasks in Automatic Driving System, because object detection is necessary in identifying and localizing cars and pedestrians. However, the accuracy and inference speed are still far away from satisfying, and the problem of missed detection occurs frequently. To address the problems, this paper first uniformly scales all dimensions of depth, width, and resolution in feature extraction module, and then fuses the semantic information in bidirectional cross-scale connections, which extract more significant features and detect objects of various scales. To decrease the undetected rate of small objects, this paper further introduces an IoU-based center localization confidence, which predicts centerness and the distance between the center of predicted box and the center of the ground-truth box, so that the predicted point approaches the center of an object. The experimental results on Cityscapes show that the proposed detector brings an improvement of 6.5% Bbox mAP and reaches an inference speed of 36.6 FPS. Compared with the other widely-used detectors such as Cascade R-CNN, FCOS, and YOLOX, it has obvious advantages in both precision and real-time performance. With additional ticks, the detection precision can be further improved by 4.4%. The proposed detector can accurately and quickly identify and localize objects, which is beneficial for the safety of self-driving cars.

IVP-YOLOv5: an intelligent vehicle-pedestrian detection method based on YOLOv5s

Computer vision is now vital in intelligent vehicle environment perception systems. However, real-time small-scale pedestrian detection in intelligent vehicle environment perception systems is still needs to be improved. This paper proposes an intelligent vehicle-pedestrian detection method based on YOLOv5s, named IVP-YOLOv5, to use in vehicle environment perception systems. Based on the network structure of YOLOv5s, we replaced BottleNeck CSP with Ghost-Bottleneck to reduce the complexity of processing feature maps while maintaining good detection performance. To reduce the error between the ground truth box and the predicted box, we apply Alpha-IoU as the bounding box loss function, improving pedestrian detection accuracy and robustness. We introduce the slicing-aided hyper inference (SAHI) strategy, which enables the lightweight backbone network to capture more detailed features of pedestrians by enlarging image pixels. Experiments on the BDD100 K dataset show that the proposed IVP-YOLOv5 achieves 67.1% AP and 18.5% APs of pedestrian detection, and the GFLOPs and the number of parameters are only 10.5 and 4.9M.

KCFS-YOLOv5: A High-Precision Detection Method for Object Detection in Aerial Remote Sensing Images

Aerial remote sensing image object detection, based on deep learning, is of great significance in geological resource exploration, urban traffic management, and military strategic information. To improve intractable problems in aerial remote sensing image, we propose a high-precision object detection method based on YOLOv5 for aerial remote sensing image. The object detection method is called KCFS-YOLOv5. To obtain the appropriate anchor box, we used the K-means++ algorithm to optimize the initial clustering points. To further enhance the feature extraction and fusion ability of the backbone network, we embedded the Coordinate Attention (CA) in the backbone network of YOLOv5 and introduced the Bidirectional Feature Pyramid Network (BiFPN) in the neck network of conventional YOLOv5. To improve the detection precision of tiny objects, we added a new tiny object detection head based on the conventional YOLOv5. To reduce the deviation between the predicted box and the ground truth box, we used the SIoU Loss function. Finally, we fused and adjusted the above improvement points and obtained high-precision detection method: KCFS-YOLOv5. This detection method was evaluated on three datasets (NWPU VHR-10, RSOD, and UCAS-AOD-CAR). The comparative experiment results demonstrate that our KCFS-YOLOv5 has the highest accuracy for the object detection in aerial remote sensing image.

3-D HANet: A Flexible 3-D Heatmap Auxiliary Network for Object Detection

3D object detection is a vital part of outdoor scene perception. Learning the complete size and accurate positioning of objects from an incomplete point cloud spatial structure is essential to 3D object detection. We propose a novel flexible 3D heatmap auxiliary network for object detection (3D HANet). To obtain complete structure and location information from an incomplete point cloud structure, we propose a 3D heatmap to reflect object information. Also, we design a plug-and-play auxiliary network based on 3D heatmap, which improves the accuracy of the entire detection network without extra computation in the inference stage. We validate the 3D heatmap auxiliary network on the basis of three classic 3D object detection networks: PointPillars, SECOND and SASSD. Experimental results show that our auxiliary network augments the feature extraction ability of the backbone network, which is manifested in that the predicted boxes and the ground truth boxes are more suitable in size and more aligned in direction. Furthermore, we conducted verification experiments on the state-of-the-art detector, CasA, and made a further improvement on the official ranking of KITTI dataset.

FI-WSOD: Foreground Information Guided Weakly Supervised Object Detection

Existing solutions for weakly supervised object detection (WSOD) generally follow the multiple instance learning (MIL) paradigm to formulate WSOD as a multi-class classification problem over a set of region proposals. However, without the supervision signal of ground-truth boxes, the training objective of multi-class classification makes the detectors devote main efforts to finding the most common pattern of each class, as the common pattern is always the most discriminative evidence for classification. In addition, although learning from distinguishing multiple foreground classes, the detectors can still ignore to differentiate foreground regions from the background ones, which causes false alarm in prediction. These two points account for the limited localization capability of MIL-based WSOD methods. To this end, we propose foreground information guided WSOD (FI-WSOD), a novel framework that introduces an extra foreground-background binary classification (F-BBC) sub-task to the original MIL-based WSOD paradigm. Particularly, the proposed FI-WSOD ameliorates object detectors at both training and inference stages. At the training stage, the involvement of F-BBC task not only improves the feature representation of the network, but also provides extra information from the foreground-background perspective. By leveraging the learnt foreground information, a Foreground Guided Self-Training (FGST) module is further proposed to filter out noisy samples, and to mine representative seeds from the remaining proposals. Moreover, a Multi-Seed Training strategy is performed to reduce the impact of noisy labels when training the self-training networks in FGST. At the inference stage, the F-BBC results are utilized to update the initial multi-class classification scores, which are further integrated with FGST results for better performance. We have conducted extensive experiments on the prevalent Pascal VOC 2007, Pascal VOC 2012 and MSCOCO datasets, and report a series of state-of-the-art records achieved by our proposed framework.

WSRC: Weakly Supervised Faster RCNN Toward Accurate Traffic Object Detection

Weakly supervised object detection has attracted increasing interest in recent years as driven by data annotation bottleneck in deep network training. On the other hand, the performance of weakly supervised models still lags far behind due to the missing of object location information during training. This paper presents a Weakly Supervised Faster RCNN (WSRC) that aims to train accurate traffic object detectors without using any box annotations, addressing the phenomenon that the traffic object detector trained by traditional method cannot be applied in the real traffic scene due to the lack of data annotation. The WSRC designs an innovative proposal evaluation network (PEN) that is trained by using a large amount of proposals that are scored according to their Intersection-over-Union (IoU) overlap with pseudo ground-truth boxes. Novel loss functions are designed to teach the PEN and WSRC to identify good proposals and regress to accurate object boxes under the guidance of loss dynamics and proposal scores. The proposed WSRC achieves superior detection performance over the datasets Pascal VOC2007, MS COCO and Cityscape, demonstrating its effectiveness while short of annotation boxes.

Concrete crack detection using lightweight attention feature fusion single shot multibox detector

As one of the most important defects of concrete, cracks seriously threaten the service life and safety of concrete structures, and various safety incidents caused by the collapse of concrete structures have occurred. Therefore, it is essential to detect concrete cracks as soon as possible. Existing object detection methods have low detection accuracy for cracks, leading to unsatisfactory detection results. In this paper, we propose a variety of feasible modules that improve the accuracy of single shot multibox detection (SSD), which is the most efficient object detection method in terms of both accuracy and speed. First, to improve the neural network’s ability to learn high-level and low-level feature maps, we propose a feature fusion enhancement module (FFEM). Second, to more effectively capture the information between feature map channels, we propose convolutional network attention (CNA). Third, to improve the anchor box fit to the ground truth box, we reset the distribution of the anchor box. Last, we propose a new type of nonmaximum suppression (NMS) named T-Soft NMS to address numerous issues with current NMS and to significantly enhance the performance of the model. We tested our method on a crack dataset, and numerous tests showed that it outperformed competing methods. In addition, we carried out ablation studies to confirm the validity and efficacy of our method.

Teacher–Student Behavior Recognition in Classroom Teaching Based on Improved YOLO-v4 and Internet of Things Technology

Based on the classroom teaching scenarios, an improved YOLO-v4 behavior detection algorithm is proposed to recognize the behaviors of teachers and students. With the development of CNN (Convolutional Neural Networks) and IoT (Internet of Things) technologies, target detection algorithms based on deep learning have become mainstream, and typical algorithms such as SSD (Single Shot Detection) and YOLO series have emerged. Based on the videos or images collected in the perception layer of the IoT paradigm, deep learning models are used in the processing layer to implement various intelligent applications. However, none of these deep learning-based algorithms are perfect, and there is room for improvement in terms of detection accuracy, computing speed, and multi-target detection capabilities. In this paper, by introducing the concept of cross-stage local network, embedded connection (EC) components are constructed and embedded at the end of the YOLO-v4 network to obtain an improved YOLO-v4 network. Aiming at the problem that it is difficult to quickly and effectively identify the students’ actions when they are occluded, the Repulsion loss function is connected in series on the basis of the original YOLO-v4 loss function. The newly added loss function consists of two parts: RepGT loss and RepBox loss. The RepGT loss function is used to calculate the loss values between the target prediction box and the adjacent ground truth boxes to reduce false positive detection results; the RepBox loss function is used to calculate the loss value between the target prediction box and other adjacent target prediction boxes to reduce false negative detection results. The training and testing are carried out on the classroom behavior datasets of teachers and students, respectively. The experimental results show that the average precision of identifying various classroom behaviors of different targets exceeds 90%, which verifies the effectiveness of the proposed method. The model performs well in sustainable classroom behavior recognition in educational context, accurate recognition of classroom behaviors can help teachers and students better understand classroom learning and promote the development of intelligent classroom model.

BANet: Small and multi-object detection with a bidirectional attention network for traffic scenes

Improving the detection accuracy and speed for small and multi-object detection is a hot issue in traffic environments. Despite the substantial advances in object detection algorithms based on deep neural networks, addressing the inaccuracy and low efficiency of small and multi-object detection remains challenging. In this paper, we propose a bidirectional attention network called BANet, which includes multichannel attention (MCA) blocks, alpha-effective intersection-over-union (α-EIoU) loss, and a multiple attention fusion (MAF) module. Each MCA block consists of low-layer, medium-layer, and high-layer features to provide rich base information for feature fusion at the neck module. We introduce MAF to alleviate the spatial location loss and poor semantic performance resulting from the continuous downsampling of the path aggregation feature pyramid network (PAFPNet). Finally, α-EIoU is our regression loss module, which calculates the difference between the predicted box and the ground truth (gt) box. Our study further demonstrates that these strategies yield significant improvements in performance over some existing YOLO detectors. Compared with the performance of YOLOX, BANet demonstrates 0.39%–0.52% mAP@0.5 improvement on the PASCAL VOC 2007 (VOC 07) dataset and 0.55%–2.93% mAP@0.5 improvement on the PASCAL VOC 2012 (VOC 12) dataset. Additionally, 0.3%–1.01% improvement in the mAP@0.5 is achieved on the MS COCO 2017 (COCO 17) dataset, indicating that BANet has a significant effect on multi-object detection. Experiments to determine the approximate number of parameters with YOLOX, show that our strategy not only improves by 7.5 frames per second (FPS) but also reduces the Average forward time by 0.97 ms.

A Detection Model for Cucumber Root-Knot Nematodes Based on Modified YOLOv5-CMS

The development of resistant cucumber varieties is of a great importance for reducing the production loss caused by root-knot nematodes. After cucumber plants are infected with root-knot nematodes, their roots will swell into spherical bumps. Rapid and accurate detection of the infected sites and assessment of the disease severity play a key role in selecting resistant cucumber varieties. Because the locations and sizes of the spherical bumps formed after different degrees of infection are random, the currently available detection and counting methods based on manual operation are extremely time-consuming and labor-intensive, and are prone to human error. In response to these problems, this paper proposes a cucumber root-knot nematode detection model based on the modified YOLOv5s model (i.e., YOLOv5-CMS) in order to support the breeding of resistant cucumber varieties. In the proposed model, the dual attention module (CBAM-CA) was adopted to enhance the model’s ability of extracting key features, the K-means++ clustering algorithm was applied to optimize the selection of the initial cluster center, which effectively improved the model’s performance, and a novel bounding box regression loss function (SIoU) was used to fuse the direction information between the ground-truth box and the predicted box so as to improve the detection precision. The experiment results show that the recall (R) and mAP of the YOLOv5s-CMS model were improved by 3% and 3.1%, respectively, compared to the original YOLOv5s model, which means it can achieve a better performance in cucumber root-knot nematode detection. This study provides an effective method for obtaining more intuitive and accurate data sources during the breeding of cucumber varieties resistant to root-knot nematode.

Visual Navigation Algorithm for Night Landing of Fixed-Wing Unmanned Aerial Vehicle

In the recent years, visual navigation has been considered an effective mechanism for achieving an autonomous landing of Unmanned Aerial Vehicles (UAVs). Nevertheless, with the limitations of visual cameras, the effectiveness of visual algorithms is significantly limited by lighting conditions. Therefore, a novel vision-based autonomous landing navigation scheme is proposed for night-time autonomous landing of fixed-wing UAV. Firstly, due to the difficulty of detecting the runway caused by the low-light image, a strategy of visible and infrared image fusion is adopted. The objective functions of the fused and visible image, and the fused and infrared image, are established. Then, the fusion problem is transformed into the optimal situation of the objective function, and the optimal solution is realized by gradient descent schemes to obtain the fused image. Secondly, to improve the performance of detecting the runway from the enhanced image, a runway detection algorithm based on an improved Faster region-based convolutional neural network (Faster R-CNN) is proposed. The runway ground-truth box of the dataset is statistically analyzed, and the size and number of anchors in line with the runway detection background are redesigned based on the analysis results. Finally, a relative attitude and position estimation method for the UAV with respect to the landing runway is proposed. New coordinate reference systems are established, six landing parameters, such as three attitude and three positions, are further calculated by Orthogonal Iteration (OI). Simulation results reveal that the proposed algorithm can achieve 1.85% improvement of AP on runway detection, and the reprojection error of rotation and translation for pose estimation are 0.675∘ and 0.581%, respectively.

DEFINING THE FUNCTIONAL SPACE OF BUS STOPS FROM MLS POINT CLOUDS

Abstract. Urban transports are essential for the mobility of citizens, therefore, functional spaces allocated to them should be correctly defined. This research defines the functional spaces for bus stops based on Mobile Laser Scanning (MLS) point clouds. Specifically, three functional spaces are defined; functional space allocated for the marquee (pedestrian space), functional space allocated for the road markings (bus space), and functional space allocated for the bus stop (pedestrians and bus). The method presented is divided into four stages; 1) Bus stop location in the point cloud from given geographical coordinates, 2) Selection of the region of interest (ROI), 3) Classification of the elements in the ROI, 4) Functional space definition. Several algorithms are used to reduce the computational time (point density reduction) and to classify the different elements contained in the bus stop (DBSCAN, height, width, and intensity filter). In addition, bounding boxes are defined to delimit the functional spaces, being compared with the ground truth box. Seven bus stops sited in Palencia (Spain) were analysed. Results show an F1-score value of 0.98 regarding the classification of the elements. Functional spaces allocated for marquees were perfectly defined. Although most of the estimated functional spaces allocated for the bus and the bus stop have high accuracy, in some cases this accuracy decreases due to the functional space allocated for the bus was overestimated.

CBASH: Combined Backbone and Advanced Selection Heads With Object Semantic Proposals for Weakly Supervised Object Detection

Most recent object detection methods have achieved growing performance on public datasets. However, enormous efforts are needed for these methods due to the extensive annotations of ground-truth boxes. Weakly Supervised Object Detection (WSOD) methods hence have been proposed to solve this problem as only image-level annotations are required and then output bounding boxes related to the objects. In order to further elevate the weakly supervised detection methods on the extraction of reasonable features, the training of potential positive proposals, and the generation of proposals before training, we propose a new Combined Backbone and Advanced Selection Heads (CBASH) method with the proposals generated from the object semantic information. Specifically, Combined Backbone will make the unobvious object features more noticeable, Advanced Selection Heads promote more potential positive proposals to get training, and the generated object semantic proposals elevate the quality and quantity of positive proposals. The proposed method is evaluated on the challenging PASCAL VOC 2007 and 2012 benchmark datasets. Experimental results show that our proposed method can achieve improved performance on both VOC 2007 and VOC 2012 datasets and outperforms the existing state-of-the-art methods.

Depth-Guided Progressive Network for Object Detection

Multi-scale object detection in natural scenes is still challenging. To enhance the multi-scale perception capability, some algorithms combine the lower-level and higher-level information via multi-scale feature fusion strategies. However, the inherent spatial properties among instances and relations between foreground and background are ignored. In addition, the human-defined “center-based” regression quality evaluation strategy, predicting a high-to-low score based on a linear relationship with the distance to the center of ground-truth box, is not robust to scale-variant objects. In this work, we propose a Depth-Guided Progressive Network (DGPNet) for multi-scale object detection. Specifically, besides the prediction of classification and localization, the depth is estimated and used to guide the image features in a weighted manner to obtain a better spatial representation. Therefore, depth estimation and 2D object detection are simultaneously learned via a unified network, where the depth features are merged as auxiliary information into the detection branch to enhance the discrimination among multi-scale objects. Moreover, to overcome the difficulty of empirically fitting the localization quality function, high-quality predicted boxes on scale-variant objects are more adaptively obtained by an IoU-aware progressive sampling strategy. We divide the sampling process into two stages, i.e., “statistical-aware” and “IoU-aware”. The former selects thresholds for positive samples based on statistical characteristics of multi-scale instances, and the latter further selects high-quality samples by IoU on the basis of the former. Therefore, the final ranking scores better reflect the quality of localization. Experiments verify that our method outperforms state-of-the-art methods on the KINS and Cityscapes dataset.