Average Precision Research Articles

Lightweight and accurate aphid detection model based on an improved deep-learning network

Rapid and accurate detection of bamboo aphids can help prevent large-scale aphid infestations from occurring, which is of great significance for increasing bamboo shoot production and economic benefits. Herein, a lightweight and accurate model, SCA-YOLOv5s, was established by integrating ShuffleNetv2 and Coordinate Attention with the YOLOv5s model to detect Takecallis taiwanus on the yellow sticky traps. Specifically, we first replaced the backbone network of YOLOv5s with ShuffleNetv2 to reduce the number of parameters and computational complexity of the model. Second, an anchor optimization method was proposed by combining linear scaling and k-means algorithm to generate appropriate anchor boxes for detecting small-sized alate aphids. Third, the coordinate attention mechanism was added to the neck network to improve the feature extraction ability. To verify the performance of the proposed SCA-YOLOv5s model, eight detection models were constructed with existing deep learning methods, including SSD300, YOLOv3, Faster R-CNN, YOLOv4, YOLOv4-Tiny, YOLOX-Tiny, YOLOv7-Tiny, and YOLOv5s. Results reveal that the SCA-YOLOv5s model achieved higher detection accuracy than the other eight models. Its mean average precision reached 92.2 %. The proposed model has a size of only 6.7 MB, its floating-point operations (FLOPs) is 7.4 × 109, its inference time is 6.6 ms, and compared with YOLOv5s, it is 53.47 % smaller in model size, 55.15 % lower in FLOPs, and 0.8 ms faster in inference time. The results indicate that the proposed model can maintain high detection accuracy while minimizing computation and inference time, which is crucial for deployment in remote areas with low information technology. This study provides valuable technical support for the control of aphids in bamboo forests.

A lightweight feature extraction backbone for gangue detection with improved Resnet50

ABSTRACT In response to the sizable parameters and floating point operations (FLOPs) of feature extraction backbone of gangue detection algorithm, a lightweight backbone was proposed by improving Resnet50. Initially, the partial 3 × 3 convolution was used to replace regular 3 × 3 convolution in original Resnet50 for reducing parameters and FLOPs in convolution operation. Subsequently, layers of batch normalization and rectified linear unit after 3 × 3 convolution were removed for their further reduction. Finally, bottleneck attention module was integrated at the bottleneck of Resnet50 to improve the mean average precision (mAP) in gangue detection. Based on three proposed methods, the ablation and comparative experiments were carried out. The reduction in parameters and FLOPs and increase in mAP were analyzed and discussed according to experimental results. It could be concluded that the parameters and FLOPs of the obtained lightweight backbone were no more than 56.36% and 57.42% of those of the popular backbones, while the decrease amplitude of its mAP value was no more than 0.2%.

Abnormal Detection of Commutator Surface Defects Based on YOLOv8

The YOLOv8 model has high detection efficiency and classification accuracy in detecting commutator surface defects, aimed at the problem of low working efficiency of a commutator, caused by commutator surface defects. First, the theoretical framework of Region-based Convolutional Neural Networks (R-CNN), spatial pyramid pooling (SPP)-net, Fast R-CNN, and Faster R-CNN is introduced, and the detection principle and process are described in detail. Secondly, the principle of the YOLOv8 network structure, head structure, neck structure, and C2f module are explained, and the loss function is described. The average precision of the proposed algorithm for detecting cracks and small points is more than 98%, and the frames per second (FPS) is 27. The detection results are mapped to the original image, and the visualization of the commutator surface defect detection is obtained, which has a higher robustness, accuracy, and real-time performance than the R-CNN, SPP-net, Fast R-CNN, and Faster R-CNN algorithms.

Yolo Based Defects Detection Algorithm for EL in PV Modules with Focal and Efficient IoU Loss

Considering the defect detection issues in electroluminescence (EL) of photovoltaic (PV) cell systems, lots of factors result in performance degradation, including defect diversity, data imbalance, scale difference, etc. Focal-EIoU loss, an effective defect detection solution for EL, is proposed based on the improved YOLOv5. Firstly, by analyzing the detection background and scale characteristics of EL defects, a binary classification is carried out in the system. Subsequently, a cascade detection network based on YOLOv5 is designed to further extract features from the binary-classified defects. The defect localization and classification are achieved in this way. To address the problem of imbalanced defect samples, a loss function is designed based on EIoU and Focal-F1 Loss. Experimental results are illustrated to show the effectiveness. Compared with the existing CNN-based deep learning approaches, the proposed focal loss calculation-based method can effectively improve the performance of handling sample imbalance. Moreover, in the detection of 12 types of defects, the Yolov5 algorithms can always obtain higher MAP (mean average precision) even with different parameter levels (Yolov5m: 0.791 vs. 0.857, Yolov5l: 0.798 vs. 0.862, Yolov5x: 0.802 vs. 0.867, Yolov5s: 0.793 vs. 0.865).

An efficient explainable deep neural network classifier for diabetic retinopathy detection

Early detection of Diabetic Retinopathy (DR) is crucial for effective intervention, particularly given its often asymptomatic nature in initial stages. Automated detection systems offer significant advantages, such as improving screening efficiency, extending healthcare accessibility to remote regions, and facilitating proactive disease management. Introducing an innovative framework for retinal image analysis that combines explainable deep learning, hybrid feature extraction, and advanced data augmentation to optimize performance and interpretability for clinical applications. The xDNN model, trained on the MESSIDOR-2 dataset, achieved an average recall of 98%, precision of 98.2%, F1-score of 98%, and accuracy of 98.2%. When extensively trained on the APTOS 2019 dataset, the model delivered outstanding results with an average precision, recall, and F1-score of 99%, and an accuracy of 99.7%. Additionally, the model's performance on the IDRID dataset was remarkable, with average precision, recall, F1-score, and accuracy all reaching 99%. Noteworthy is our method impressive average Area Under the Curve (AUC) of 99.8%, affirming its consistent and exceptional performance across all classes of diabetic retinopathy. This underscores the xDNN Classifier's potential as a valuable tool for precise and reliable diabetic retinopathy detection. It holds substantial promise for elevating clinical diagnosis and enhancing treatment outcomes within the field of ophthalmology.

Multivariable model for gait pattern differentiation in elderly patients with hip and knee osteoarthritis: A wearable sensor approach

BackgroundHip and knee osteoarthritis (OA) patients demonstrate distinct gait patterns, yet detecting subtle abnormalities with wearable sensors remains uncertain. This study aimed to assess a predictive model's efficacy in distinguishing between hip and knee OA gait patterns using accelerometer data. MethodParticipants with hip or knee OA underwent overground walking assessments, recording lower limb accelerations for subsequent time and frequency domain analyses. Logistic regression with regularization identified associations between frequency domain features of acceleration signals and OA, and k-nearest neighbor classification distinguished knee and hip OA based on selected acceleration signal features. FindingsWe included 57 knee OA patients (30 females, median age 68 [range 49–89], median BMI 29.7 [range 21.0–45.9]) and 42 hip OA patients (19 females, median age 70 [range 47–89], median BMI 28.3 [range 20.4–37.2]). No significant difference could be found in the time domain's averaged shape of acceleration signals. However, in the frequency domain, five selected features showed a diagnostic ability to differentiate between knee and hip OA. Using these features, a model achieved a 77 % accuracy in classifying gait cycles into hip or knee OA groups, with average precision, recall, and F1 score of 77 %, 76 %, and 78 %, respectively. InterpretationThe study demonstrates the effectiveness of wearable sensors in differentiating gait patterns between individuals with hip and knee OA, specifically in the frequency domain. The results highlights the promising potential of wearable sensors and advanced signal processing techniques for objective assessment of OA in clinical settings.

Diabetic Retinopathy Features Segmentation without Coding Experience with Computer Vision Models YOLOv8 and YOLOv9.

Computer vision is a powerful tool in medical image analysis, supporting the early detection and classification of eye diseases. Diabetic retinopathy (DR), a severe eye disease secondary to diabetes, accompanies several early signs of eye-threatening conditions, such as microaneurysms (MAs), hemorrhages (HEMOs), and exudates (EXs), which have been widely studied and targeted as objects to be detected by computer vision models. In this work, we tested the performances of the state-of-the-art YOLOv8 and YOLOv9 architectures on DR fundus features segmentation without coding experience or a programming background. We took one hundred DR images from the public MESSIDOR database, manually labelled and prepared them for pixel segmentation, and tested the detection abilities of different model variants. We increased the diversity of the training sample by data augmentation, including tiling, flipping, and rotating the fundus images. The proposed approaches reached an acceptable mean average precision (mAP) in detecting DR lesions such as MA, HEMO, and EX, as well as a hallmark of the posterior pole of the eye, such as the optic disc. We compared our results with related works in the literature involving different neural networks. Our results are promising, but far from being ready for implementation into clinical practice. Accurate lesion detection is mandatory to ensure early and correct diagnoses. Future works will investigate lesion detection further, especially MA segmentation, with improved extraction techniques, image pre-processing, and standardized datasets.

Development and validation of a perivascular space segmentation method in multi-center datasets

BackgroundPerivascular spaces (PVS) visible on magnetic resonance imaging (MRI) are significant markers associated with various neurological diseases. Although quantitative analysis of PVS may enhance sensitivity and improve consistency across studies, the field lacks a universally validated method for analyzing images from multi-center studies. MethodsWe annotated PVS on multi-center 3D T1-weighted (T1w) images acquired using scanners from three major vendors (Siemens, General Electric, and Philips). A neural network, mcPVS-Net (multi-center PVS segmentation network), was trained using data from 40 subjects and then tested in a separate cohort of 15 subjects. We assessed segmentation accuracy against ground truth masks tailored for each scanner vendor. Additionally, we evaluated the agreement between segmented PVS volumes and visual scores for each scanner. We also explored correlations between PVS volumes and various clinical factors such as age, hypertension, and white matter hyperintensities (WMH) in a larger sample of 1020 subjects. Furthermore, mcPVS-Net was applied to a new dataset comprising both T1w and T2-weighted (T2w) images from a United Imaging scanner to investigate if PVS volumes could discriminate between subjects with differing visual scores. We also compared the mcPVS-Net with a previously published method that segments PVS from T1 images. ResultsIn the test dataset, mcPVS-Net achieved a mean DICE coefficient of 0.80, with an average Precision of 0.81 and Recall of 0.79, indicating good specificity and sensitivity. The segmented PVS volumes were significantly associated with visual scores in both the basal ganglia (r = 0.541, p < 0.001) and white matter regions (r = 0.706, p < 0.001), and PVS volumes were significantly different among subjects with varying visual scores. Segmentation performance was consistent across different scanner vendors. PVS volumes exhibited significant associations with age, hypertension, and WMH. In the United Imaging scanner dataset, PVS volumes showed good associations with PVS visual scores evaluated on either T1w or T2w images. Compared to a previously published method, mcPVS-Net showed a higher accuracy and improved PVS segmentation in the basal ganglia region. ConclusionThe mcPVS-Net demonstrated good accuracy for segmenting PVS from 3D T1w images. It may serve as a useful tool for future PVS research.

An Enhanced SL-YOLOv8-Based Lightweight Remote Sensing Detection Algorithm for Identifying Broken Strands in Transmission Lines

Power transmission lines frequently face threats from lightning strikes, severe storms, and chemical corrosion, which can lead to damage in steel–aluminum-stranded wires, thereby seriously affecting the stability of the power system. Currently, manual inspections are relatively inefficient and high risk, while drone inspections are often limited by complex environments and obstacles. Existing detection algorithms still face difficulties in identifying broken strands. To address these issues, this paper proposes a new method called SL-YOLOv8. This method incorporates an improved You Only Look Once version 8 (YOLOv8) algorithm, specifically designed for online intelligent inspection robots to detect broken strands in transmission lines. Transmission lines are susceptible to lightning strikes, storms, and chemical corrosion, which is leading to the potential failure of steel- and aluminum-stranded lines, and significantly impacting the stability of the power system. Currently, manual inspections come with relatively low efficiency and high risk, and Unmanned Aerial Vehicle (UAV) inspections are hindered by complex situations and obstacles, with current algorithms making it difficult to detect the broken strand lines. This paper proposes SL-YOLOv8, which is a broken transmission line strand detection method for an online intelligent inspection robot combined with an improved You Only Look Once version 8 (YOLOv8). By incorporating the Squeeze-and-Excitation Network version 2 (SENet_v2) into the feature fusion network, the method effectively enhances adaptive feature representation by focusing on and amplifying key information, thereby improving the network’s capability to detect small objects. Additionally, the introduction of the LSKblockAttention module, which combines Large Selective Kernels (LSKs) and the attention mechanism, allows the model to dynamically select and enhance critical features, significantly enhancing detection accuracy and robustness while maintaining model precision. Compared with the original YOLOv8 algorithm, SL-YOLOv8 demonstrates improved precision recognition accuracy in Break-ID-1632 and cable damage datasets. The precision is increased by 3.9% and 2.7%, and the recall is increased by 12.2% and 2.3%, respectively, for the two datasets. The mean average precision (mAP) at the Intersection over Union (IoU) threshold of 0.5 is also increased by 4.9% and 1.2%, showing the SL-YOLOv8’s effectiveness in accurately identifying small objects in complex situations.

Leaf rolling detection in maize under complex environments using an improved deep learning method

Leaf rolling is a common adaptive response that plants have evolved to counteract the detrimental effects of various environmental stresses. Gaining insight into the mechanisms underlying leaf rolling alterations presents researchers with a unique opportunity to enhance stress tolerance in crops exhibiting leaf rolling, such as maize. In order to achieve a more profound understanding of leaf rolling, it is imperative to ascertain the occurrence and extent of this phenotype. While traditional manual leaf rolling detection is slow and laborious, research into high-throughput methods for detecting leaf rolling within our investigation scope remains limited. In this study, we present an approach for detecting leaf rolling in maize using the YOLOv8 model. Our method, LRD-YOLO, integrates two significant improvements: a Convolutional Block Attention Module to augment feature extraction capabilities, and a Deformable ConvNets v2 to enhance adaptability to changes in target shape and scale. Through experiments on a dataset encompassing severe occlusion, variations in leaf scale and shape, and complex background scenarios, our approach achieves an impressive mean average precision of 81.6%, surpassing current state-of-the-art methods. Furthermore, the LRD-YOLO model demands only 8.0 G floating point operations and the parameters of 3.48 M. We have proposed an innovative method for leaf rolling detection in maize, and experimental outcomes showcase the efficacy of LRD-YOLO in precisely detecting leaf rolling in complex scenarios while maintaining real-time inference speed.

Centralized feature pyramid‐based supervised deep learning for object detection model from GPR data

AbstractTo address low detection accuracy and speed due to the multisolvability of the ground‐penetrating radar signal, we proposed a novel centralized feature pyramid‐YOLOv6l–based model to enhance detection precision and speed in road damage and pipeline detection. The centralized feature pyramid was used to obtain rich intra‐layer features and improve the network performance. Our proposed model achieves higher accuracy compared with the existing detection models. We also built two new evaluating indexes, relative average precision and relative mean average precision, to fully evaluate the detection accuracy. To verify the applicability of our model, we conducted a road field detection experiment on a ground‐penetrating radar dataset we collected and found that the proposed model had good performance in increasing detection precision, achieving the highest mean average precision compared with YOLOv7, YOLOv5 and YOLOx models, with relative mean average precision and frame rate per second at 16.38% and 30.5%, respectively. The detection information for the road damage and pipeline were used to conduct three‐dimensional imaging. Our model is suitable for object detection in ground‐penetrating radar images, thereby providing technical support for road damage and underground pipeline detection.

DART: An automated end-to-end object detection pipeline with data Diversification, open-vocabulary bounding box Annotation, pseudo-label Review, and model Training

Accurate real-time object detection is vital across numerous industrial applications, from safety monitoring to quality control. Traditional approaches, however, are hindered by arduous manual annotation and data collection, struggling to adapt to ever-changing environments and novel target objects. To address these limitations, this paper presents DART, an innovative automated end-to-end pipeline that revolutionizes object detection workflows from data collection to model evaluation. It eliminates the need for laborious human labeling and extensive data collection while achieving outstanding accuracy across diverse scenarios. DART encompasses four key stages: (1) Data Diversification using subject-driven image generation (DreamBooth with SDXL), (2) Annotation via open-vocabulary object detection (Grounding DINO) to generate bounding box and class labels, (3) Review of generated images and pseudo-labels by large multimodal models (InternVL-1.5 and GPT-4o) to guarantee credibility, and (4) Training of real-time object detectors (YOLOv8 and YOLOv10) using the verified data. We apply DART to a self-collected dataset of construction machines named Liebherr Product, which contains over 15K high-quality images across 23 categories. The current instantiation of DART significantly increases average precision (AP) from 0.064 to 0.832. Its modular design ensures easy exchangeability and extensibility, allowing for future algorithm upgrades, seamless integration of new object categories, and adaptability to customized environments without manual labeling and additional data collection. The code and dataset are released at https://github.com/chen-xin-94/DART.

Radar-Based Elderly Fall Detection Using Smoothed Pseudo Wigner Ville Distribution and XGBoost Learning.

Falls among the elderly population pose significant health risks, often leading to morbidity and decreased quality of life. Traditional fall detection methods, namely wearable devices and cameras, have limitations such as lighting conditions and privacy concerns. Radar-based fall detection has emerged as a promising alternative, offering unobtrusive technique. In this study, an attempt has been made to classify fall detection using smoothed pseudo wigner-ville distribution (SPWVD) images and XGBoost learning. For this, online publicly available radar database (N=15) is considered. Radar signals is employed to SPWVD for time-frequency representation images. Ten features are extracted and applied to XGBoost learning. Experiments are performed and performance is evaluated using 10-fold cross validation. The proposed approach is able to discriminate elderly fall. Using XGBoost learning, the approach yields a maximum average classification accuracy, f1-score, precision, sensitivity, specificity, and kappa scores of 87.47%, 87.38%, 88.12%, 86.81%, 88.31% and 74.94% respectively. The combination of conventional features with concentration measures and median frequency obtained the second best performance. Thus, the proposed framework could be utilized for accurate and efficient detection of falls among the elderly population in their private spaces.

Human Multi-Activities Classification Using mmWave Radar: Feature Fusion in Time-Domain and PCANet.

This study introduces an innovative approach by incorporating statistical offset features, range profiles, time-frequency analyses, and azimuth-range-time characteristics to effectively identify various human daily activities. Our technique utilizes nine feature vectors consisting of six statistical offset features and three principal component analysis network (PCANet) fusion attributes. These statistical offset features are derived from combined elevation and azimuth data, considering their spatial angle relationships. The fusion attributes are generated through concurrent 1D networks using CNN-BiLSTM. The process begins with the temporal fusion of 3D range-azimuth-time data, followed by PCANet integration. Subsequently, a conventional classification model is employed to categorize a range of actions. Our methodology was tested with 21,000 samples across fourteen categories of human daily activities, demonstrating the effectiveness of our proposed solution. The experimental outcomes highlight the superior robustness of our method, particularly when using the Margenau-Hill Spectrogram for time-frequency analysis. When employing a random forest classifier, our approach outperformed other classifiers in terms of classification efficacy, achieving an average sensitivity, precision, F1, specificity, and accuracy of 98.25%, 98.25%, 98.25%, 99.87%, and 99.75%, respectively.

Sugarcane disease recognition through visible and near-infrared spectroscopy using deep learning assisted continuous wavelet transform-based spectrogram

Utilizing visible and near-infrared (Vis-NIR) spectroscopy in conjunction with chemometrics methods has been widespread for identifying plant diseases. However, a key obstacle involves the extraction of relevant spectral characteristics. This study aimed to enhance sugarcane disease recognition by combining convolutional neural network (CNN) with continuous wavelet transform (CWT) spectrograms for spectral features extraction within the Vis-NIR spectra (380–1400 nm) to improve the accuracy of sugarcane diseases recognition. Using 130 sugarcane leaf samples, the obtained one-dimensional CWT coefficients from Vis-NIR spectra were transformed into two-dimensional spectrograms. Employing CNN, spectrogram features were extracted and incorporated into decision tree, K-nearest neighbour, partial least squares discriminant analysis, and random forest (RF) calibration models. The RF model, integrating spectrogram-derived features, demonstrated the best performance with an average precision of 0.9111, sensitivity of 0.9733, specificity of 0.9791, and accuracy of 0.9487. This study may offer a non-destructive, rapid, and accurate means to detect sugarcane diseases, enabling farmers to receive timely and actionable insights on the crops’ health, thus minimizing crop loss and optimizing yields.

Towards an open pipeline for the detection of critical infrastructure from satellite imagery—a case study on electrical substations in The Netherlands

Abstract Critical infrastructure (CI) are at risk of failure due to the increased frequency and magnitude of climate extremes related to climate change. It is thus essential to include them in a risk management framework to identify risk hotspots, develop risk management policies and support adaptation strategies to enhance their resilience. However, the lack of information on the exposure of CI to natural hazards prevents their incorporation in large-scale risk assessments. This study sets out to improve the representation of CI for risk assessment studies by building a neural network model to detect CI assets from optical remote sensing imagery. We present a pipeline that extracts CI from OpenStreetMap (OSM), processes the imagery and assets’ masks, and trains a Mask R-CNN model that allows for instance segmentation of CI at the asset level. This study provides an overview of the pipeline and tests it with the detection of electrical substations assets in the Netherlands. Several experiments are presented for different under-sampling percentages of the majority class (25 %, 50 % and 100 %) and hyperparameters settings (batch size and learning rate). The highest scoring experiment achieved an Average Precision at an Intersection over Union of 50 % of 30.93 and a tile F-score of 89.88 %. This allows us to confirm the feasibility of the method and invite disaster risk researchers to use this pipeline for other infrastructure types. We conclude by exploring the different avenues to improve the pipeline by addressing the class imbalance, Transfer Learning and Explainable Artificial Intelligence.

Segmentation prompts classification: A nnUNet-based 3D transfer learning framework with ROI tokenization and cross-task attention for esophageal cancer T-stage diagnosis

The computer-aided diagnosis system for esophageal cancer (EC) holds vital significance in EC diagnosis and treatment making, with a primary focus on accurate segmentation of EC-related organs and classification of EC’s T-stage. Above two tasks are closely related and crucial in assisting surgeon segment and diagnose cancer early. Note that this paradigm is still at its infancy and limited by closely related open issues: (1) how to link the complementary relationship between these two tasks and improve the originally poor performance? and (2) how to determine whether the tumor has invaded the surrounding muscle layers from CT images? Aiming at these issues, this study develops nn-TransEC, a 3D transfer learning framework that builds upon nnU-Net and synergizes segmentation and classification. nn-TransEC focuses on prompting fine-grained classification of EC’s T-stage with the aid of prior segmentation, which is implemented in two parts: (1) A nnUNet-configured multi-task learning network (nn-MTNet) is designed for complementary segmentation of EC-related organs and classification of EC’s T-stages with cross-task attention gates and transfer learning. (2) A knowledge-embedded ROI tokenization method (KRT) is defined to mimic the diagnostic workflow of doctors for classifying EC’s T-stage. KRT is implemented by cropping the most concerned regions from entire CT volume based on prior segmentation. Experiments have been conducted on a private dataset collected from 169 patients with confirmed EC through pathological diagnosis. Our proposed nn-TransEC is compared against the state-of-the-art counterparts (e.g., nnU-Net and nnFormer), and results demonstrate that: nn-TransEC excels in all compared methods in multi-organ segmentation and classification of EC’s T-stages, with 3D Dice of EC and average AUC of T-stages reaching 0.844 and 0.941, respectively. In contrast, the state-of-the-art method nnFormer achieves 0.814 and 0.927, respectively. Meanwhile, nn-TransEC also outperforms state-of-the-art multi-task learning models in joint segmentation and classification, with Hausdorff Distance of EC and average precision of T-stages reaching 8.497 and 0.845, respectively. In contrast, the state-of-the-art method TransMT-Net achieves 12.206 and 0.730, respectively.

Enhancing Pulmonary Embolism Detection in COVID-19 Patients Through Advanced Deep Learning Techniques.

The intersection of COVID-19 and pulmonary embolism (PE) has posed unprecedented challenges in medical diagnostics. The critical nature of PE and its increased incidence during the pandemic underline the need for improved detection methods. This study evaluates the effectiveness of advanced deep learning techniques in enhancing PE detection in post-COVID-19 patients through Computed Tomography Pulmonary Angiography (CTPA) scans. Using a dataset of 746 anonymized CTPA images from 25 patients, we fine-tuned the state-of-the-art Ultralytics YOLOv8 object detection model, which was trained on 676 images with 1,517 annotated bounding boxes and validated on 70 images with 108 bounding boxes. After 200 epochs of training, which lasted approximately 1.021 hours, the YOLOv8 model demonstrated significant diagnostic proficiency, achieving a mean Average Precision (mAP) of 0.683 at an IoU threshold of 0.50 and a mAP of 0.246 at the IoU range of 0.50:0.95 in the validation dataset. Notably, the model reached a maximum precision of 0.85949 and a maximum recall of 0.81481, though these metrics were observed in separate epochs. These findings emphasize the model's potential for high diagnostic accuracy and offer a promising direction for deploying AI tools in clinical settings, significantly contributing to healthcare innovation and patient care post-pandemic.

Roof plane parsing towards LoD-2.2 building reconstruction based on joint learning using remote sensing images

Building models are important for urban studies. Remote sensing multi-spectral (MS) images are widely used for its rich semantic information. The lack of geometry features is fulfilled by introducing photogrammetry derived digital surface models (DSMs), resulting in pairs of DSMs and MS images. Utilizing such pairs and a convolutional neural network, level of detail (LoD) 2.2 building models, which contain roof planes and major roof elements (e.g. dormers), are reconstructed in this work. Leveraging both raster and vector predictions, 3-D building models with straight edges and sharp corners are obtained. The proposed two-stage method first extracts vectorized roof lines from pairs of DSMs and RGB images, followed by generation of detailed 2-D and 3-D polygonal building models. We conducted our experiments based on two datasets: a custom dataset in Landsberg am Lech in Germany, and an open dataset named Roof3D. For the custom dataset, our proposed model achieved mean average precision (mAP) for building roof vertices of 64.3% and for building roof lines of 54.5% at highest. Mean precision and recall for reconstructed 2-D building roof plane polygons are 52.2% and 54.7% respectively. For the Roof3D dataset, mAP is reported to be 25.3% and 12.4% for the extracted building roof lines and roof plane polygons respectively.

Diffusion Models-Based Purification for Common Corruptions on Robust 3D Object Detection.

LiDAR sensors have been shown to generate data with various common corruptions, which seriously affect their applications in 3D vision tasks, particularly object detection. At the same time, it has been demonstrated that traditional defense strategies, including adversarial training, are prone to suffering from gradient confusion during training. Moreover, they can only improve their robustness against specific types of data corruption. In this work, we propose LiDARPure, which leverages the powerful generation ability of diffusion models to purify corruption in the LiDAR scene data. By dividing the entire scene into voxels to facilitate the processes of diffusion and reverse diffusion, LiDARPure overcomes challenges induced from adversarial training, such as sparse point clouds in large-scale LiDAR data and gradient confusion. In addition, we utilize the latent geometric features of a scene as a condition to assist the generation of diffusion models. Detailed experiments show that LiDARPure can effectively purify 19 common types of LiDAR data corruption. Further evaluation results demonstrate that it can improve the average precision of 3D object detectors to an extent of 20% in the face of data corruption, much higher than existing defence strategies.