Bounding Box Research Articles

Greater benefits of deep learning-based computer-aided detection systems for finding small signals in 3D volumetric medical images.

Radiologists are tasked with visually scrutinizing large amounts of data produced by 3D volumetric imaging modalities. Small signals can go unnoticed during the 3D search because they are hard to detect in the visual periphery. Recent advances in machine learning and computer vision have led to effective computer-aided detection (CADe) support systems with the potential to mitigate perceptual errors. Sixteen nonexpert observers searched through digital breast tomosynthesis (DBT) phantoms and single cross-sectional slices of the DBT phantoms. The 3D/2D searches occurred with and without a convolutional neural network (CNN)-based CADe support system. The model provided observers with bounding boxes superimposed on the image stimuli while they looked for a small microcalcification signal and a large mass signal. Eye gaze positions were recorded and correlated with changes in the area under the ROC curve (AUC). The CNN-CADe improved the 3D search for the small microcalcification signal ( , ) and the 2D search for the large mass signal ( , ). The CNN-CADe benefit in 3D for the small signal was markedly greater than in 2D ( , ). Analysis of individual differences suggests that those who explored the least with eye movements benefited the most from the CNN-CADe ( , ). However, for the large signal, the 2D benefit was not significantly greater than the 3D benefit ( , ). The CNN-CADe brings unique performance benefits to the 3D (versus 2D) search of small signals by reducing errors caused by the underexploration of the volumetric data.

Lightweight-detection: The strip steel surface defect identification based on improved YOLOv5d

The surface defect of strip steel seriously affects the product quality, the current identification methods have the inherent defects of low detection efficiency and poor detection accuracy, and it is important an improved an algorithm based on the modified YOLOv5 framework to enhance the defect detection effectiveness and accuracy. To create a more lightweight network model, the C3 module and part of the convolutional structure in the original YOLOv5 network were replaced with the GhostBottleneck structure. Additionally, the Coordinate Attention (CA) mechanism was incorporated into the Backbone section to compensate for the original model's lack of attention mechanisms. To address the angular mismatch between the actual frame and the predicted bounding box, a limitation overlooked by the CIOU loss function was further refined. Experimental results demonstrate that the enhanced YOLOv5s-CSG model outperformed the original YOLOv5s by 7.3 %, achieving a significant 37.9 % reduction in model size and an mAP value of 77.5 %. Furthermore, the detection precision and speed were notably higher compared to several widely-used algorithms. The proposed YOLOv5s-CSG model is well-suited for deployment on mobile devices and shows great potential for rapidly and accurately detecting both the type and location of strip surface defects, surpassing the performance of existing methods.

A method for assessing guide layout and error on implant accuracy

Functional and esthetic results require accurate implant placement. We aimed to develop a predictive method for assessing guide layout and error on implant accuracy. A mathematical model for position error analysis was constructed based on triangular mesh data. This model examines the relationship between the spatial shifts of multiple surfaces and the spatial shifts of specific points. It involves encasing these surfaces in a cuboid bounding box and expressing them in a local coordinate system. The influence of positioning surface error and layout of surgical guide were researched with a simulation test. The result shows that error in the implant site position is directly related to the error in the guide locating surface under the same layout. When the guide locating surface layout varies, as the length, width, and height of the minimum cuboid envelope increase, the maximum deviation in the implant site position decreases.

Simulation method of asphalt mixture based on fracturable digital aggregate model

This study proposed an approach to construct a fracturable aggregate model to make the fracture patterns consistent with the real situation. Firstly, this study utilized three-dimensional structured light technology to obtain the geometric shell of aggregate and divide them into four categories of spherical, angular, elongated and flaky aggregates according to the minimum bounding box algorithm. Secondly, this study proposed a discrete-spheres filling algorithm based on the Bubble Pack Filling algorithm to fill the shell accurately. Thirdly, this study established the conversion equations between the macro and micro parameters of aggregate by using the linear parallel bonding model, and used it to establish the fracturable aggregate model. Through the indoor and virtual static fracture tests, this study found that the fracture patterns of the fracturable aggregate models are similar to the real situation and characterized the fracture patterns of four types of aggregates. Further analyzing the virtual static fracture tests, this study found that the bearing capacity of aggregates is determined by the compressive strength of host rocks and the shape of aggregate. Through the indoor and virtual uniaxial compression creep tests, this study investigated the behavior of the mixture specimens consisted by the fracturable aggregate model. The stress-strain curves of the fracturable and unfracturable aggregate model in this test were plotted in this study, and the results showed that the relative error between the results obtained by the fracturable model and those of the indoor test is much smaller than that of the unfracturable model. Considering that existing simulation tests on asphalt mixture often ignore aggregate deformation and fracture, so this study used the method mentioned above to obtain the fracturable aggregate model, which can improve the confidence of asphalt mixture simulation results.

YOLOv8 Model for Weed Detection in Wheat Fields Based on a Visual Converter and Multi-Scale Feature Fusion.

Accurate weed detection is essential for the precise control of weeds in wheat fields, but weeds and wheat are sheltered from each other, and there is no clear size specification, making it difficult to accurately detect weeds in wheat. To achieve the precise identification of weeds, wheat weed datasets were constructed, and a wheat field weed detection model, YOLOv8-MBM, based on improved YOLOv8s, was proposed. In this study, a lightweight visual converter (MobileViTv3) was introduced into the C2f module to enhance the detection accuracy of the model by integrating input, local (CNN), and global (ViT) features. Secondly, a bidirectional feature pyramid network (BiFPN) was introduced to enhance the performance of multi-scale feature fusion. Furthermore, to address the weak generalization and slow convergence speed of the CIoU loss function for detection tasks, the bounding box regression loss function (MPDIOU) was used instead of the CIoU loss function to improve the convergence speed of the model and further enhance the detection performance. Finally, the model performance was tested on the wheat weed datasets. The experiments show that the YOLOv8-MBM proposed in this paper is superior to Fast R-CNN, YOLOv3, YOLOv4-tiny, YOLOv5s, YOLOv7, YOLOv9, and other mainstream models in regards to detection performance. The accuracy of the improved model reaches 92.7%. Compared with the original YOLOv8s model, the precision, recall, mAP1, and mAP2 are increased by 10.6%, 8.9%, 9.7%, and 9.3%, respectively. In summary, the YOLOv8-MBM model successfully meets the requirements for accurate weed detection in wheat fields.

Detection of Cervical Lesion Cell/Clumps Based on Adaptive Feature Extraction.

Automated detection of cervical lesion cell/clumps in cervical cytological images is essential for computer-aided diagnosis. In this task, the shape and size of the lesion cell/clumps appeared to vary considerably, reducing the detection performance of cervical lesion cell/clumps. To address the issue, we propose an adaptive feature extraction network for cervical lesion cell/clumps detection, called AFE-Net. Specifically, we propose the adaptive module to acquire the features of cervical lesion cell/clumps, while introducing the global bias mechanism to acquire the global average information, aiming at combining the adaptive features with the global information to improve the representation of the target features in the model, and thus enhance the detection performance of the model. Furthermore, we analyze the results of the popular bounding box loss on the model and propose the new bounding box loss tendency-IoU (TIoU). Finally, the network achieves the mean Average Precision (mAP) of 64.8% on the CDetector dataset, with 30.7 million parameters. Compared with YOLOv7 of 62.6% and 34.8M, the model improved mAP by 2.2% and reduced the number of parameters by 11.8%.

Advanced techniques for automated emotion recognition in dogs from video data through deep learning

Inter-species emotional relationships, particularly the symbiotic interaction between humans and dogs, are complex and intriguing. Humans and dogs share fundamental mammalian neural mechanisms including mirror neurons, crucial to empathy and social behavior. Mirror neurons are activated during the execution and observation of actions, indicating inherent connections in social dynamics across species despite variations in emotional expression. This study explores the feasibility of using deep-learning Artificial Intelligence systems to accurately recognize canine emotions in general environments, to assist individuals without specialized knowledge or skills in discerning dog behavior, particularly related to aggression or friendliness. Starting with identifying key challenges in classifying pleasant and unpleasant emotions in dogs, we tested advanced deep-learning techniques and aggregated results to distinguish potentially dangerous human--dog interactions. Knowledge transfer is used to fine-tune different networks, and results are compared on original and transformed sets of frames from the Dog Clips dataset to investigate whether DogFACS action codes detailing relevant dog movements can aid the emotion recognition task. Elaborating on challenges and biases, we emphasize the need for bias mitigation to optimize performance, including different image preprocessing strategies for noise mitigation in dog recognition (i.e., face bounding boxes, segmentation of the face or body, isolating the dog on a white background, blurring the original background). Systematic experimental results demonstrate the system’s capability to accurately detect emotions and effectively identify dangerous situations or signs of discomfort in the presence of humans.

Accurately Computing the Interacted Volume of Molecules over Their 3D Mesh Models.

For quickly predicting the rational arrangement of catalysts and substrates, we previously proposed a method to calculate the interacted volumes of molecules over their 3D point cloud models. However, the nonuniform density in molecular point clouds may lead to incomplete contours in some slices, reducing the accuracy of the previous method. In this paper, we propose a two-step method for more accurately computing molecular interacted volumes. First, by employing a prematched mesh slicing method, we layer the 3D triangular mesh models of the electrostatic potential isosurfaces of two molecules globally, transforming the volume calculation into finding the intersecting areas in each layer. Next, by subdividing polygonal edges, we accurately identify intersecting parts within each layer, ensuring precise calculation of interacted volumes. In addition, we present a concise overview for computing intersecting areas in cases of multiple contour intersections and for improving computational efficiency by incorporating bounding boxes at three stages. Experimental results demonstrate that our method maintains high accuracy in different experimental data sets, with an average relative error of 0.16%. On the same experimental setup, our average relative error is 0.07%, which is lower than the previous algorithm's 1.73%, improving the accuracy and stability in calculating interacted volumes.

A 2.5D Self-Training Strategy for Carotid Artery Segmentation in T1-Weighted Brain Magnetic Resonance Images.

Precise annotations for large medical image datasets can be time-consuming. Additionally, when dealing with volumetric regions of interest, it is typical to apply segmentation techniques on 2D slices, compromising important information for accurately segmenting 3D structures. This study presents a deep learning pipeline that simultaneously tackles both challenges. Firstly, to streamline the annotation process, we employ a semi-automatic segmentation approach using bounding boxes as masks, which is less time-consuming than pixel-level delineation. Subsequently, recursive self-training is utilized to enhance annotation quality. Finally, a 2.5D segmentation technique is adopted, wherein a slice of a volumetric image is segmented using a pseudo-RGB image. The pipeline was applied to segment the carotid artery tree in T1-weighted brain magnetic resonance images. Utilizing 42 volumetric non-contrast T1-weighted brain scans from four datasets, we delineated bounding boxes around the carotid arteries in the axial slices. Pseudo-RGB images were generated from these slices, and recursive segmentation was conducted using a Res-Unet-based neural network architecture. The model's performance was tested on a separate dataset, with ground truth annotations provided by a radiologist. After recursive training, we achieved an Intersection over Union (IoU) score of (0.68 ± 0.08) on the unseen dataset, demonstrating commendable qualitative results.

Quantification of cardiac capillarization in basement-membrane-immunostained myocardial slices using Segment Anything Model

Decreased myocardial capillary density has been reported as an important histopathological feature associated with various heart disorders. Quantitative assessment of cardiac capillarization typically involves double immunostaining of cardiomyocytes (CMs) and capillaries in myocardial slices. In contrast, single immunostaining of basement membrane protein is a straightforward approach to simultaneously label CMs and capillaries, presenting fewer challenges in background staining. However, subsequent image analysis always requires expertise and laborious manual work to identify and segment CMs/capillaries. Here, we developed an image analysis tool, AutoQC, for automatic identification and segmentation of CMs and capillaries in immunofluorescence images of basement membrane. Commonly used capillarization-related measurements can be derived from segmentation results. By leveraging the power of a pre-trained segmentation model (Segment Anything Model, SAM) via prompt engineering, the training of AutoQC required only a small dataset with bounding box annotations instead of pixel-wise annotations. AutoQC outperformed SAM (without prompt engineering) and YOLOv8-Seg, a state-of-the-art instance segmentation model, in both instance segmentation and capillarization assessment. Thus, AutoQC, featuring a weakly supervised algorithm, enables automatic segmentation and high-throughput, high-accuracy capillarization assessment in basement-membrane-immunostained myocardial slices. This approach reduces the training workload and eliminates the need for manual image analysis once AutoQC is trained.

BS-YOLOV8: an intelligent detection model for bearing pin support-piece states of high-rise building machine

Abstract As the main support part of the working platform of a high-rise building machine, the bearing pin support (BPS) plays a crucial role in the safety and stability of the platform, the conventional method has the problems of low detection efficiency, low accuracy, and high cost. To improve the accuracy and robustness of the detection algorithm under weak light, this paper proposes an intelligent detection algorithm for the BPS-piece states of the BS-YOLOV8, to improve the feature map utilization and reduce the model leakage detection error detection rate, Swin transformer is used to improve the YOLOV8 backbone network. In addition, the BiFormer attention mechanism is used to weigh the feature map to solve the problem of feature information loss in different feature layers and weak lighting conditions, and then the Scylla-IOU loss function is used instead of the original localization loss function to guide the model to learn to generate a predicted bounding box closer to the real target bounding box. Finally, the BS-YOLOV8 algorithm is used to compare with its classical algorithm on the self-constructed dataset of this study, The results show that the mAP0.5, mAP0.5:0.95, and FPS values of the BS-YOLOV8 algorithm reach 97.9%, 96.3% and 40 under normal lighting. The mAP0.5 value reaches 87.6% under low light conditions, which effectively solves the problems of low detection efficiency and poor detection under low light conditions, and is superior compared to other algorithms.

EMG-YOLO: road crack detection algorithm for edge computing devices.

Road cracks significantly shorten the service life of roads. Manual detection methods are inefficient and costly. The YOLOv5 model has made some progress in road crack detection. However, issues arise when deployed on edge computing devices. The main problem is that edge computing devices are directly connected to sensors. This results in the collection of noisy, poor-quality data. This problem adds computational burden to the model, potentially impacting its accuracy. To address these issues, this paper proposes a novel road crack detection algorithm named EMG-YOLO. First, an Efficient Decoupled Header is introduced in YOLOv5 to optimize the head structure. This approach separates the classification task from the localization task. Each task can then focus on learning its most relevant features. This significantly reduces the model's computational resources and time. It also achieves faster convergence rates. Second, the IOU loss function in the model is upgraded to the MPDIOU loss function. This function works by minimizing the top-left and bottom-right point distances between the predicted bounding box and the actual labeled bounding box. The MPDIOU loss function addresses the complex computation and high computational burden of the current YOLOv5 model. Finally, the GCC3 module replaces the traditional convolution. It performs global context modeling with the input feature map to obtain global context information. This enhances the model's detection capabilities on edge computing devices. Experimental results show that the improved model has better performance in all parameter indicators compared to current mainstream algorithms. The EMG-YOLO model improves the accuracy of the YOLOv5 model by 2.7%. The mAP (0.5) and mAP (0.9) are improved by 2.9% and 0.9%, respectively. The new algorithm also outperforms the YOLOv5 model in complex environments on edge computing devices. The EMG-YOLO algorithm proposed in this paper effectively addresses the issues of poor data quality and high computational burden on edge computing devices. This is achieved through optimizing the model head structure, upgrading the loss function, and introducing global context modeling. Experimental results demonstrate significant improvements in both accuracy and efficiency, especially in complex environments. Future research can further optimize this algorithm and explore more lightweight and efficient object detection models for edge computing devices.

A streamlined approach for intelligent ship object detection using EL-YOLO algorithm

Maritime objects frequently exhibit low-quality and insufficient feature information, particularly in complex maritime environments characterized by challenges such as small objects, waves, and reflections. This situation poses significant challenges to the development of reliable object detection including the strategies of loss function and the feature understanding capabilities in common YOLOv8 (You Only Look Once) detectors. Furthermore, the widespread adoption and unmanned operation of intelligent ships have generated increasing demands on the computational efficiency and cost of object detection hardware, necessitating the development of more lightweight network architectures. This study proposes the EL-YOLO (Efficient Lightweight You Only Look Once) algorithm based on YOLOv8, designed specifically for intelligent ship object detection. EL-YOLO incorporates novel features, including adequate wise IoU (AWIoU) for improved bounding box regression, shortcut multi-fuse neck (SMFN) for a comprehensive analysis of features, and greedy-driven filter pruning (GDFP) to achieve a streamlined and lightweight network design. The findings of this study demonstrate notable advancements in both detection accuracy and lightweight characteristics across diverse maritime scenarios. EL-YOLO exhibits superior performance in intelligent ship object detection using RGB cameras, showcasing a significant improvement compared to standard YOLOv8 models.

The Effectiveness of Sensor Visualizations and Graphic Augmentations for Detecting Vertical Obstacles

Slow or failed detection of low‐salience vertical obstacles and associated wires is one of today’s leading causes of fatal helicopter accidents. The risk of collisions with such obstacles is likely to increase as advanced aerial mobility and broadening drone activity promises to increase the density of air traffic at low altitudes, while growing demand for electricity and communication will expand the number of vertical structures. The current see‐and‐avoid detection paradigm relies on pilots to spend much of their visual attention looking outside for obstacles. This method is inadequate in low‐visibility conditions, cluttered environments, and given the need for pilots to engage in multiple competing visual tasks. With the expected growing number of hazards and an increased traffic volume, the current approach to collision avoidance will become even less tenable. A human‐in‐the‐loop helicopter simulator study was conducted to assess the effectiveness of sensor visualizations (image intensification or thermal imaging) and graphic augmentations (a bounding box around a tower and a circle surrounding the base of the tower) for supporting fast and reliable detection of vertical structures. Graphic augmentations resulted in faster tower detection time when ambient visibility and illumination were reduced close to the limit for visual flight. Bounding boxes around towers were detected first in all conditions but tended to mask the obstacle they were meant to highlight. Sensor visualization affected tower detection time only at night, where night vision goggles were more effective than the infrared thermal sensor.

Box2Mask: Box-Supervised Instance Segmentation via Level-Set Evolution.

In contrast to fully supervised methods using pixel-wise mask labels, box-supervised instance segmentation takes advantage of simple box annotations, which has recently attracted increasing research attention. This paper presents a novel single-shot instance segmentation approach, namely Box2Mask, which integrates the classical level-set evolution model into deep neural network learning to achieve accurate mask prediction with only bounding box supervision. Specifically, both the input image and its deep features are employed to evolve the level-set curves implicitly, and a local consistency module based on a pixel affinity kernel is used to mine the local context and spatial relations. Two types of single-stage frameworks, i.e., CNN-based and transformer-based frameworks, are developed to empower the level-set evolution for box-supervised instance segmentation, and each framework consists of three essential components: instance-aware decoder, box-level matching assignment and level-set evolution. By minimizing the level-set energy function, the mask map of each instance can be iteratively optimized within its bounding box annotation. The experimental results on five challenging testbeds, covering general scenes, remote sensing, medical and scene text images, demonstrate the outstanding performance of our proposed Box2Mask approach for box-supervised instance segmentation. In particular, with the Swin-Transformer large backbone, our Box2Mask obtains 42.4% mask AP on COCO, which is on par with the recently developed fully mask-supervised methods.

LSOTB-TIR: A Large-Scale High-Diversity Thermal Infrared Single Object Tracking Benchmark.

Unlike visual object tracking, thermal infrared (TIR) object tracking methods can track the target of interest in poor visibility such as rain, snow, and fog, or even in total darkness. This feature brings a wide range of application prospects for TIR object-tracking methods. However, this field lacks a unified and large-scale training and evaluation benchmark, which has severely hindered its development. To this end, we present a large-scale and high-diversity unified TIR single object tracking benchmark, called LSOTB-TIR, which consists of a tracking evaluation dataset and a general training dataset with a total of 1416 TIR sequences and more than 643 K frames. We annotate the bounding box of objects in every frame of all sequences and generate over 770 K bounding boxes in total. To the best of our knowledge, LSOTB-TIR is the largest and most diverse TIR object tracking benchmark to date. We spilt the evaluation dataset into a short-term tracking subset and a long-term tracking subset to evaluate trackers using different paradigms. What's more, to evaluate a tracker on different attributes, we also define four scenario attributes and 12 challenge attributes in the short-term tracking evaluation subset. By releasing LSOTB-TIR, we encourage the community to develop deep learning-based TIR trackers and evaluate them fairly and comprehensively. We evaluate and analyze 40 trackers on LSOTB-TIR to provide a series of baselines and give some insights and future research directions in TIR object tracking. Furthermore, we retrain several representative deep trackers on LSOTB-TIR, and their results demonstrate that the proposed training dataset significantly improves the performance of deep TIR trackers. Codes and dataset are available at https://github.com/QiaoLiuHit/LSOTB-TIR.

The End-to-End Fetal Head Circumference Detection and Estimation in Ultrasound Images.

In prenatal examinations, the fetal head circumference (HC) measurement is essential for assessing fetal weight and health conditions. The sonographers obtain the fetal HC manually by fitting peripheral skull ellipse in clinical practice, which is highly subjective, time-consuming, and experience-dependent. Recently, many fetal HC automatic measurement algorithms have been proposed to improve workflow efficiency in prenatal examination. But most automatic measurement algorithms focus on using fetal head segmentation as an intermediate processing step, and HC estimation relies heavily on segmentation results, which causes the accumulation of errors in the above two stages. Independent of the segmentation method, we design a regression network to generate the oriented bounding box to detect the head contour, and directly obtain the fetal head parameters with a pixel-based ellipse regression (PER) loss. Moreover, an effective 3D attention mechanism is integrated into the network to estimate HC more precisely without adding parameters in complex ultrasound images. The extensive experimental results on the public HC18 and our clinical dataset show that the proposed network provides a feasible scheme for end-to-end estimating fetal HC, and avoids the mistake brought by the intermediary processes.

A multiscale 3D network for lung nodule detection using flexible nodule modeling.

Lung cancer is the most common type of cancer. Detection of lung cancer at an early stage can reduce mortality rates. Pulmonary nodules may represent early cancer and can be identified through computed tomography (CT) scans. Malignant risk can be estimated based on attributes like size, shape, location, and density. Deep learning algorithms have achieved remarkable advancements in this domain compared to traditional machine learning methods. Nevertheless, many existing anchor-based deep learning algorithms exhibit sensitivity to predefined anchor-box configurations, necessitating manual adjustments to obtain optimal outcomes. Conversely, current anchor-free deep learning-based nodule detection methods normally adopt fixed-size nodule models like cubes or spheres. To address these technical challenges, we propose a multiscale 3D anchor-free deep learning network (M3N) for pulmonary nodule detection, leveraging adjustable nodule modeling (ANM). Within this framework, ANM empowers the representation of target objects in an anisotropic manner, with a novel point selection strategy (PSS) devised to accelerate the learning process of anisotropic representation. We further incorporate a composite loss function that combines the conventional L2 loss and cosine similarity loss, facilitating M3N to learn nodules' intensity distribution in three dimensions. Experiment results show that the M3N achieves 90.6% competitive performance metrics (CPM) with seven predefined false positives per scan on the LUNA 16 dataset. This performance appears to exceed that of other state-of-the-art deep learning-based networks reported in their respective publications. Individual test results also demonstrate that M3N excels in providing more accurate, adaptive bounding boxes surrounding the contours of target nodules. The newly developed nodule detection system reduces reliance on prior knowledge, such as the general size of objects in the dataset, thus it should enhance overall robustness and versatility. Distinct from traditional nodule modeling techniques, the ANM approach aligns more closely with the morphological characteristics of nodules. Time consumption and detection results demonstrate promising efficiency and accuracy which should be validated in clinical settings.

A CNN-KNN Based Recognition of Online Handwritten Symbols within Physics Expressions Using Contour-Based Bounding Box (CBBS) Segmentation Technique

A CNN-KNN Based Recognition of Online Handwritten Symbols within Physics Expressions Using Contour-Based Bounding Box (CBBS) Segmentation Technique

Spatio-Temporal Contextual Learning for Single Object Tracking on Point Clouds.

Single object tracking (SOT) is one of the most active research directions in the field of computer vision. Compared with the 2-D image-based SOT which has already been well-studied, SOT on 3-D point clouds is a relatively emerging research field. In this article, a novel approach, namely, the contextual-aware tracker (CAT), is investigated to achieve a superior 3-D SOT through spatially and temporally contextual learning from the LiDAR sequence. More precisely, in contrast to the previous 3-D SOT methods merely exploiting point clouds in the target bounding box as the template, CAT generates templates by adaptively including the surroundings outside the target box to use available ambient cues. This template generation strategy is more effective and rational than the previous area-fixed one, especially when the object has only a small number of points. Moreover, it is deduced that LiDAR point clouds in 3-D scenes are often incomplete and significantly vary from frame to another, which makes the learning process more difficult. To this end, a novel cross-frame aggregation (CFA) module is proposed to enhance the feature representation of the template by aggregating the features from a historical reference frame. Leveraging such schemes enables CAT to achieve a robust performance, even in the case of extremely sparse point clouds. The experiments confirm that the proposed CAT outperforms the state-of-the-art methods on both the KITTI and NuScenes benchmarks, achieving 3.9% and 5.6% improvements in terms of precision.