Single-stage Detector Research Articles

YOLO vs CNN : A Comparative Study on Object Detection

Abstract—With the availability of large amount of data and increasing demand of digitizing the visual systems, lot of research on object detection techniques has been the foremost requirement since last ten years. As the technological advancements are pacing at such a higher rate with increasing computational power and new architectures for CNN has already been proposed in a successive manner to enhance the detection techniques. For Real time Object identification, CNN has proposed several architectures but YOLO proves itself better than others based on the speed of identification. This paper presents the brief overview and comparative analysis of various object detection techniques like SSD, RCNN, YOLO and their subsequent versions for different applications. On comparing, it is found that single stage detectors outperform two stage detectors. Keywords—Pascal VOC, MS COCO, YOLO, CNN, Object detection, Object Recognition, SSD.

Dq-YOLOF: An Effective Improvement with Deformable Convolution and Sample Quality Optimization Based on the YOLOF Detector

Single-stage detectors have drawbacks of insufficient accuracy and poor coverage capability. YOLOF (You Only Look One-level Feature) has achieved better performance in this regard, but there is still room for improvement. To enhance the coverage capability for objects of different scales, we propose an improved single-stage object detector: Dq-YOLOF. We have designed an output encoder that employs a series of modules utilizing deformable convolution and SimAM (Simple Attention Module). This module replaces the dilated convolution in YOLOF. This design significantly improves the ability to express details. Simultaneously, we have redefined the sample selection strategy, which optimizes the quality of positive samples based on SimOTA. It can dynamically allocate positive samples according to their quality, reducing computational load and making it more suitable for small objects. Experiments conducted on the COCO 2017 dataset also verify the effectiveness of our method. Dq-YOLOF achieved 38.7 AP, 1.5 AP higher than YOLOF. To confirm performance improvements on small objects, our method was tested on urinary sediment and aerial drone datasets for generalization. Notably, it enhances performance while also lowering computational costs.

Weakly Supervised Underwater Object Real-time Detection Based on High-resolution Attention Class Activation Mapping and Category Hierarchy

Recently, deep learning-based underwater object detection technology has achieved remarkable success. However, the accuracy and completeness of dataset instance annotation are crucial for its success. The quality of underwater images is low, severe objects clustering, and occlusion, acquiring object's annotations demands substantial time and labor costs, while mis annotation and missed annotation can also degrade model performance and limit their application in practical scenarios. To address this issue, this paper presents a novel weakly supervised underwater object real-time detection method, which is divided into two subtasks: weakly supervised object localization and real-time object detection. In the weakly supervised object localization task, we design a novel category hierarchy structure network that integrates the high-resolution attention-class activation mapping algorithm to obtain high-quality object class activation maps, weaken background interference, and obtain more complete object regions. The parameterized spatial loss module is devised to enable the model to escape from local optimal solutions, thus accurately and efficiently obtaining object pseudo-detection annotation boxes. For the real-time object detection task, the single-stage detector YOLOv7 is selected as the basic detection model, and an object perception loss function is designed based on the class activation map to jointly supervise the training process. A method for filtering noisy pseudo-supervision information is proposed to enhance the pseudo-supervision information involved in training. Ablation experiments and multi-method comparison experiments were conducted on the URPC and RUOD datasets, and the results verify the effectiveness of the proposed strategy, and our model exhibits significant advantages in detection performance and detection efficiency compared to current mainstream and advanced models.

YOLO-HyperVision: A vision transformer backbone-based enhancement of YOLOv5 for detection of dynamic traffic information

With the increase of traffic flow in modern urban areas, traffic congestion has become a serious problem that affects people’s normal production and life. Using target detection technology instead of manual labor can quickly detect the road traffic situation and provide timely information about the traffic flow. However, when using drones to observe the traffic flow in the air, the perspective effect will cause the detected vehicles and pedestrians to be very small, and the scale difference between different categories of targets is large, which increases the detection difficulty of a single convolutional neural network model. In order to solve the problem of low accuracy of traditional single-stage target detection models, this study proposes an improved Yolov5 vehicle target detection model with Vision Transformer (VIT) backbone, You Only Look Once-HyperVision (YOLO-HV), which aims to solve the problem of poor multi-scale target recognition performance caused by the inability of traditional CNN networks to integrate contextual information, and help drones achieve more efficient and accurate traffic flow recognition functions. This study deeply integrates the Vision Transformer (VIT) backbone and Convolutional Neural Network (CNN), effectively combining the multi-scale detection advantages of Vision Transformer and the inductive bias ability of Convolutional Neural Network, and adds multi-scale residual modules and context correlation enhancement modules, which greatly improves the recognition accuracy of single-stage detectors for drone images. Through comparative experiments on the VisDrone dataset, it is found that the detection performance of this model is improved compared with several commonly used detection models. YOLO-HV can increase the mean average precision (mAP) by 3.3% compared with the pure convolutional network of the same model size. YOLO-HV model has achieved excellent performance in the task of traffic flow image detection taken by drones, and can more accurately identify and classify road vehicles than various target detection models.

LEPS: A Lightweight and Effective Single-Stage Detector for Pothole Segmentation

LEPS: A Lightweight and Effective Single-Stage Detector for Pothole Segmentation

Semi-Supervised One-Stage Object Detection for Maize Leaf Disease

Maize is one of the most important crops globally, and accurate diagnosis of leaf diseases is crucial for ensuring increased yields. Despite the continuous progress in computer vision technology, detecting maize leaf diseases based on deep learning still relies on a large amount of manually labeled data, and the labeling process is time-consuming and labor-intensive. Moreover, the detectors currently used for identifying maize leaf diseases have relatively low accuracy in complex experimental fields. Therefore, the proposed Agronomic Teacher, an object detection algorithm that utilizes limited labeled and abundant unlabeled data, is applied to maize leaf disease recognition. In this work, a semi-supervised object detection framework is built based on a single-stage detector, integrating the Weighted Average Pseudo-labeling Assignment (WAP) strategy and AgroYOLO detector combining Agro-Backbone network with Agro-Neck network. The WAP strategy uses weight adjustments to set objectness and classification scores as evaluation criteria for pseudo-labels reliability assignment. Agro-Backbone network accurately extracts features of maize leaf diseases and obtains richer semantic information. Agro-Neck network enhances feature fusion by utilizing multi-layer features for collaborative combinations. The effectiveness of the proposed method is validated on the MaizeData and PascalVOC datasets at different annotation ratios. Compared to the baseline model, Agronomic Teacher leverages abundant unlabeled data to achieve a 6.5% increase in mAP (0.5) on the 30% labeled MaizeData. On the 30% labeled PascalVOC dataset, the mAP (0.5) improved by 8.2%, demonstrating the method’s potential for generalization.

Efficient flexible voxel-based two-stage network for 3D object detection in autonomous driving

3D object detection from the LiDAR point cloud plays an important role in autonomous driving. It is difficult to balance inference speed and detection accuracy when performing 3D point cloud object detection due to the large size of point cloud data and its unstructured storage, which makes it difficult to represent its features. To address the challenge, we propose a two-stage point cloud object detector, AFV-RCNN. In stage-1, the attention flexible voxel feature encoding layer is introduced, which utilizes flexible voxels to enhance feature encoding speed and focuses on foreground points to be detected through voxel attention. In stage-2, the multi-level and grid-based multi-scale RoI (Region of Interest) feature fusion module is designed. It directly extracts complete 3D structures from 3D region proposals and focuses on both local and global features through multi-scale partitioning. In the training stage, GHM-C Loss is applied to address the challenges associated with imbalanced target categories and the imbalance between difficult and easy samples in the classification task. We evaluate the model on the public KITTI Dataset and Waymo Open Dataset. The mAP in KITTI for 3D detection is 73.41% and inference on a single GPU reaches 30.0 fps. Compared with other state-of-the-art methods, AFV-RCNN achieves both the inference speed of a single-stage detector and the detection accuracy of a two-stage detector. It ensures higher detection accuracy while efficiently processing the point cloud.

An adversarial example attack method based on predicted bounding box adaptive deformation in optical remote sensing images

Existing global adversarial attacks are not applicable to real-time optical remote sensing object detectors based on the YOLO series of deep neural networks, which makes it difficult to improve the adversarial robustness of single-stage detectors. The existing methods do not work well enough in optical remote sensing images, which may be due to the mechanism of adversarial perturbations is not suitable. Therefore, an adaptive deformation method (ADM) was proposed to fool the detector into generating wrong predicted bounding boxes. Building upon this, we introduce the Adaptive Deformation Method Iterative Fast Gradient Sign Method (ADM-I-FGSM) and Adaptive Deformation Mechanism Projected Gradient Descent (ADM-PGD) against YOLOv4 and YOLOv5. ADM method can obtain the deformation trend values based on the length-to-width ratio of the prediction box, and the adversarial perturbation trend generated based on these trend values has better adversarial effect. Through experiments, we validate that our approach exhibits a higher adversarial success rate compared to the state-of-the-art methods. We anticipate that our unveiled attack scheme will aid in the evaluation of adversarial resilience of these models.

Efficient Object Detection with YOLO: A Comprehensive Guide

Object detection presents itself as a pivotal and complex challenge within the domain of computer vision. Over the past ten years, as deep learning techniques have advanced quickly, researchers have committed significant resources to utilising deep models as the basis to improve the performance of object identification systems and related tasks like segmentation, localization. Two-stage and single-stage detectors are the two basic categories into which object detectors can be roughly divided. Typically, two-stage detectors use complicated structures in conjunction with a selective region proposal technique to accomplish their goals. Conversely, single-stage detectors aim to detect objects across all spatial regions in one shot, employing relatively simpler architectures. Any object detector's inference time and detection accuracy are the main factors to consider while evaluating it. Single-stage detectors offer quicker inference times, but two-stage detectors frequently show better detection accuracy. But since the introduction of YOLO (You Only Look Once) and its architectural offspring, detection accuracy has significantly improved—sometimes even outperforming that of two-stage detectors. The adoption of YOLO in various applications is primarily driven by its faster inference times rather than its detection accuracy alone.

UAV and Deep Learning: Detection of selected riparian species along the Ganga River

Abstract. Environmental protection and sustainable natural resource management are being recognised worldwide as essential goals to safe guard human health and well-being. Riparian zones, that face the highest decline in freshwater biodiversity, are of prime conservation priority because they are essential for regulating climate, preserving aquatic-terrestrial biodiversity, maintaining ground water recharge and restoring rivers. In today's fast-paced data-driven environment, artificial intelligence (AI) is the precise answer to a wide range of problems including biodiversity conservation and wildlife management. Leveraging advancements like Uncrewed/Unmanned Aerial Vehicles (UAVs) and AI has resulted innovative strides in wildlife conservation. This study utilised UAV imagery to record high-resolution data of aquatic habitat and species along the Ganga River and employed deep learning algorithms to analyse the data. Through extensive field surveys in the Hastinapur Wildlife Sanctuary, 7,025 photos representing a variety of environments, including 20,000 annotated samples of aquatic animals such as turtles and gharials were generated. Vision based computing capabilities such as pattern recognition model were developed to identify these species. To enrich and enhance the dataset and model, we used different image pre-processing techniques. Slight rotation (±5 degrees), minor cropping (up to 10%), and adjustments in brightness, saturation, and shear (±15%) were applied. Controlled blur (up to 0.5%) and exposure modifications (±5%) were also implemented on the image dataset to improve accuracy. Three Convolutional Neural Network (CNN) architectures, single-stage detectors named YOLO v7, YOLO v8, and Roboflow 3.0, were used for detecting the select species. Results show YOLO v8 excels, achieving mean average precision (mAP) of 98.8% for gharial and 92.2% for turtle detection, with a rapid average detection time of 0.308 seconds per frame at 3200 × 3200 resolution. Additionally, our model demonstrates real-time species detection capability through innovative frame sampling techniques with UAVs. This methodology provides promising technique to collect scientific data on IUCN red listed turtles and critically endangered gharials, allowing detection, monitoring, and real time counting with minimal intrusion. In conclusion, the fusion of UAVs and deep learning promises to revolutionize habitat monitoring, aiding conservation decision-making.

MsSVT++: Mixed-Scale Sparse Voxel Transformer With Center Voting for 3D Object Detection.

Accurate 3D object detection in large-scale outdoor scenes, characterized by considerable variations in object scales, necessitates features rich in both long-range and fine-grained information. While recent detectors have utilized window-based transformers to model long-range dependencies, they tend to overlook fine-grained details. To bridge this gap, we propose MsSVT++, an innovative Mixed-scale Sparse Voxel Transformer that simultaneously captures both types of information through a divide-and-conquer approach. This approach involves explicitly dividing attention heads into multiple groups, each responsible for attending to information within a specific range. The outputs of these groups are subsequently merged to obtain final mixed-scale features. To mitigate the computational complexity associated with applying a window-based transformer in 3D voxel space, we introduce a novel Chessboard Sampling strategy and implement voxel sampling and gathering operations sparsely using a hash map. Moreover, an important challenge stems from the observation that non-empty voxels are primarily located on the surface of objects, which impedes the accurate estimation of bounding boxes. To overcome this challenge, we introduce a Center Voting module that integrates newly voted voxels enriched with mixed-scale contextual information towards the centers of the objects, thereby improving precise object localization. Extensive experiments demonstrate that our single-stage detector, built upon the foundation of MsSVT++, consistently delivers exceptional performance across diverse datasets.

DEEP LEARNING FOR VISUAL RECOGNITION

The design and development of an advanced object detection system are presented in this work, which was guided by a thorough literature review and feasibility assessment. The literature review emphasises how object detection techniques have evolved, highlighting the shift from conventional techniques to deep learning approaches. Important developments are highlighted, such as feature pyramid networks, anchor-based localization, and region-based and single-stage detectors. Furthermore, offered are insights about assessment metrics, transfer learning, and data augmentation. A feasibility study assesses the suggested systems operational, technological, and financial viability and finds that it is highly feasible in each of these areas. The architecture of the system is modular and scalable, including backend services, data management, an object detection engine, and a user interface among its constituent parts. Specifications for both functional and non-functional needs are provided, which direct the development of the system. The development phases, resource allocation, development process, and quality assurance procedures are all outlined in the implementation plan. Through the integration of deep learning techniques, the suggested system seeks to achieve high-performance object identification capabilities that are appropriate for a variety of applications while being scalable, reliable, and user-friendly.

MSSD: multi-scale self-distillation for object detection

Knowledge distillation techniques have been widely used in the field of deep learning, usually by extracting valid information from a neural network with a large number of parameters and a high learning capacity (the teacher model) to a neural network with a small number of parameters and a low learning capacity (the student model). However, there are inefficiencies in the transfer of knowledge between teacher and student. The student model does not fully learn all the knowledge of the teacher model. Therefore, we aim to achieve knowledge distillation of our network layer by a single model, i.e., self-distillation. We also apply the idea of self-distillation to the object detection task and propose a multi-scale self-distillation approach, where we argue that knowledge distillation of the information contained in feature maps at different scales can help the model better detect small targets. In addition, we propose a Gaussian mask based on the target region as an auxiliary detection method to improve the accuracy of target position detection in the distillation process. We then validate our approach on the KITTI dataset using a single-stage detector YOLO. The results demonstrate a 2.8% improvement in accuracy over the baseline model without the use of a teacher model.

WS-SSD: Achieving faster 3D object detection for autonomous driving via weighted point cloud sampling

Due to the limited computational resources of the onboard computing devices of autonomous vehicles, the development of lightweight 3D object detectors is essential. Point-based detectors that progressively sample raw point clouds reduce numerous redundant computations and facilitate the implementation of high-speed 3D object detectors. The farthest point sampling based on Euclidean distance (D-FPS) is frequently utilized in point-based 3D detectors for point cloud sampling to reduce computation overhead. D-FPS can ensure uniform point sampling, covering the entire point cloud space as much as possible. The ratio of foreground and background points does not change significantly in the sampling results; hence, foreground objects do not receive sufficient attention. In addition, the number of road reflection points is large, and these points are close to foreground objects, reducing the sampling accuracy for faint objects. We propose weighted farthest point sampling based on Euclidean distance (W-DFPS), which selectively discards some road reflection points in the point sampling process, thus increasing the weight of foreground points in sampling results. It reduces the likelihood of faint objects being lost in the sampling results, such that a small number of sampling points can also cover most foreground objects. W-DFPS replaces D-FPS in the single-stage detector based on instance-aware (IA-SSD), and the network structure is slightly modified, namely the single-stage detector based on weighted sampling (WS-SSD). We evaluate WS-SSD in the KITTI dataset with a single A100 GPU. When the number of sampling points for the first module of WS-SSD is consistent with IA-SSD, the pedestrian detection accuracy is improved by an average of 4.06 compared to IA-SSD. Although the number of sampling points for the first module of WS-SSD is reduced to 25% of IA-SSD, the object detection accuracy remains competitive; it also achieves a state-of-the-art inference speed of 150.76 frames per second (FPS), which is a 46% improvement over IA-SSD.

Investigating The Relationship Between Vehicle Speed and Pothole Detection by Using Mobile Phone

It is known that road pavements are damaged due to time, climatic conditions and construction errors. Considering these damages, the most important road defect that reduces road safety and comfort is potholes. Especially as the width and depth of the pothole increases, driving safety is also endangered. In addition, the locations of these potholes, especially on urban roads, are determined manually in many regions. This process causes delays in the maintenance and repair of the potholes. To this end, the authors plan an in-vehicle integrated system consisting of multiple stages to automatically detect potholes occurring in the road network. The main purpose of the planned system is to identify potholes with high accuracy. However, the effect of vehicle speed on pothole detection in this system is unknown. In order to solve this complex situation, real-time video recordings were made on the same road and pothole at different vehicle speeds. Then, the pothole detection process was realized through these videos with the single-stage detector YOLOv7 vs YOLOv8. When the results obtained were examined, exact relationship could not be determined between vehicle speed and pothole detection. This situation may vary according to various parameters such as camera angle, image quality, sunlight condition. In addition, when both models are compared according to the performance criteria, YOLOv7 has a partial superiority over YOLOv8 in mAP0.5, precision, recall and F1 score values. It is especially significant that these criteria are close to 1. Finally, the perception results obtained from the images obtained from the video showed that there was no overfitting in the models.

Automatic detection of typical defects in reinforced concrete bridges via YOLOv5

Bridges play a crucial role in road networks, and ensuring their safety and preservation is of utmost importance for both management companies and the scientific community. Throughout their lifespan, bridges are exposed to various factors that increase their vulnerability, including aging, harsh environmental conditions, and natural hazards, all of which can potentially lead to structural failures. Following the collapse of the Polcevera Viaduct in Italy, the Ministry of Transportation proposed a comprehensive safety evaluation procedure to be implemented nationwide, aiming to develop a methodology for assessing critical cases and implementing risk mitigation measures. This process involves several stages to assign a risk class that considers different sources of hazards. Among these phases, periodic on-site surveys to identify defects and signs of degradation are required. However, several challenges arise, such as the time and cost associated with inspections, the subjectivity involved in visually identifying defects, and the need for qualified personnel. To address these issues, traditional techniques can be enhanced by leveraging digital innovations, which seek to create new and reliable tools that support road management companies in safeguarding their infrastructure assets. In this regard, deep learning-based object detectors offer promising possibilities. Specifically, automatic recognition of defects and damages on existing bridge elements can be achieved using single-stage detectors like YOLOv5. In this study, we explored the application of this technique by creating a database of typical defects and involving domain experts to label these defects. Subsequently, YOLOv5 was trained, tested, and validated, demonstrating favorable effectiveness and accuracy of the proposed methodology. This research opens new opportunities and highlights the potential of artificial intelligence in automatically detecting defects on bridges.

Benchmarking the Robustness of Instance Segmentation Models.

This article presents a comprehensive evaluation of instance segmentation models with respect to real-world image corruptions as well as out-of-domain image collections, e.g., images captured by a different set-up than the training dataset. The out-of-domain image evaluation shows the generalization capability of models, an essential aspect of real-world applications, and an extensively studied topic of domain adaptation. These presented robustness and generalization evaluations are important when designing instance segmentation models for real-world applications and picking an off-the-shelf pretrained model to directly use for the task at hand. Specifically, this benchmark study includes state-of-the-art network architectures, network backbones, normalization layers, models trained starting from scratch versus pretrained networks, and the effect of multitask training on robustness and generalization. Through this study, we gain several insights. For example, we find that group normalization (GN) enhances the robustness of networks across corruptions where the image contents stay the same but corruptions are added on top. On the other hand, batch normalization (BN) improves the generalization of the models across different datasets where statistics of image features change. We also find that single-stage detectors do not generalize well to larger image resolutions than their training size. On the other hand, multistage detectors can easily be used on images of different sizes. We hope that our comprehensive study will motivate the development of more robust and reliable instance segmentation models.

A Single Stage Detector for Breast Cancer Detection on Digital Mammogram

A Single Stage Detector for Breast Cancer Detection on Digital Mammogram

A Review and Benchmark on State-of-the-Art Steel Defects Detection

Steel, a critical material in construction, automobile, and railroad manufacturing industries, often presents defects that can lead to equipment failure, significant safety risks, and costly downtime. This research aims to evaluate the performance of state-of-the-art object detection models in detecting defects on steel surfaces, a critical task in industries such as railroad and automobile manufacturing. The study addresses the challenges of limited defect data and lengthy model training times. Five existing state-of-the-art object detection models (faster R-CNN, deformable DETR, double head R-CNN, Retinanet, and deformable convolutional network) were benchmarked on the Northeastern University (NEU) steel dataset. The selection of models covers a broad spectrum of methodologies, including two-stage detectors, single-stage detectors, transformers, and a model incorporating deformable convolutions. The deformable convolutional network achieved the highest accuracy of 77.28% on the NEU dataset following a fivefold cross-validation method. Other models also demonstrated notable performance, with accuracies within the 70–75% range. Certain models exhibited particular strengths in detecting specific defects, indicating potential areas for future research and model improvement. The findings provide a comprehensive foundation for future research in steel defect detection and have significant implications for practical applications. The research could improve quality control processes in the steel industry by automating the defect detection task, leading to safer and more reliable steel products and protecting workers by removing the human factor from hazardous environments.

A Comparative Analysis of Single-Stage Detectors from the Perspectives of Anchor-Free and Anchor-Based Approaches

Since the emergence of deep learning in object detection in 2014, there has been rapid development and widespread application of detectors across various domains such as healthcare, road traffic, and industrial inspection. The evolution of convolutional neural network (CNN)-based detectors has reached a mature stage, with single-stage detectors playing a crucial role within CNN-based frameworks. This paper provides an analysis of the developmental process and distinctive features of classic single-stage detectors categorized into two types: Anchor-free and Anchor-based. Furthermore, the paper also provides a comparison of the accuracy trends for these two types of single-stage detectors over different years. Through this analysis and comparison, potential reasons behind the advancements in single-stage detectors during different time periods can be explored, such as network architecture optimization, increase in datasets, and algorithm improvements. Additionally, this paper offers insights into current research focal points and future prospects.