Capability Of Deep Neural Networks Research Articles

Long Short-Term Memory-Based Neural Networks for Missile Maneuvers Trajectories Prediction⋆

Due to its extensive applications in different contexts, moving target tracking has become a hot topic in the last years, above all in the military field. Specifically, missile tracking research received a great effort, mainly for its importance in terms of security and safety. Herein, traditional solutions, e.g. Interacting Multiple Model (IMM) based on the Kalman estimation theory, achieve good performance under the main restrictive assumption of the a priori knowledge of the target model, so neglecting the unavoidable presence of model uncertainties and limiting the achievable tracking accuracy only by the presence of the measurement noise. With the specific aim of overcoming this narrowness, this work investigates the capability of deep neural networks in predicting the missile maneuvering trajectories in a model-free fashion. The idea is to leverage the Long-Short Term Memory (LSTM) net due to its excellent capability in learning long-term dependencies of temporal information. Two different LSTM-based architectures have been hence designed to predict both position and velocity of a missile using raw and noisy measurements provided by a realistic radar system, exploiting a large database abundant of realistic off-line data. Training results and theoretical derivations are verified through non-trivial scenarios in order to assess the capability of predicting unknown and realistic 3D missile maneuvers. Finally, the proposed approach has been also compared with a performing model-based IMM algorithm, suitably tuned to deal with realistic missile maneuvers, confirming the excellent generalization abilities of the developed data-driven architectures for different datasets.

Deep neural networks effectively model neural adaptation to changing background noise and suggest nonlinear noise filtering methods in auditory cortex

The human auditory system displays a robust capacity to adapt to sudden changes in background noise, allowing for continuous speech comprehension despite changes in background environments. However, despite comprehensive studies characterizing this ability, the computations that underly this process are not well understood. The first step towards understanding a complex system is to propose a suitable model, but the classical and easily interpreted model for the auditory system, the spectro-temporal receptive field (STRF), cannot match the nonlinear neural dynamics involved in noise adaptation. Here, we utilize a deep neural network (DNN) to model neural adaptation to noise, illustrating its effectiveness at reproducing the complex dynamics at the levels of both individual electrodes and the cortical population. By closely inspecting the model's STRF-like computations over time, we find that the model alters both the gain and shape of its receptive field when adapting to a sudden noise change. We show that the DNN model's gain changes allow it to perform adaptive gain control, while the spectro-temporal change creates noise filtering by altering the inhibitory region of the model's receptive field. Further, we find that models of electrodes in nonprimary auditory cortex also exhibit noise filtering changes in their excitatory regions, suggesting differences in noise filtering mechanisms along the cortical hierarchy. These findings demonstrate the capability of deep neural networks to model complex neural adaptation and offer new hypotheses about the computations the auditory cortex performs to enable noise-robust speech perception in real-world, dynamic environments.

Expectation-maximization algorithm leads to domain adaptation for a perineural invasion and nerve extraction task in whole slide digital pathology images.

In addition to lymphatic and vascular channels, tumor cells can also spread via nerves, i.e., perineural invasion (PNI). PNI serves as an independent prognostic indicator in many malignancies. As a result, identifying and determining the extent of PNI is an important yet extremely tedious task in surgical pathology. In this work, we present a computational approach to extract nerves and PNI from whole slide histopathology images. We make manual annotations on selected prostate cancer slides once but then apply the trained model for nerve segmentation to both prostate cancer slides and head and neck cancer slides. For the purpose of multi-domain learning/prediction and investigation on the generalization capability of deep neural network, an expectation-maximization (EM)-based domain adaptation approach is proposed to improve the segmentation performance, in particular for the head and neck cancer slides. Experiments are conducted to demonstrate the segmentation performances. The average Dice coefficient for prostate cancer slides is 0.82 and 0.79 for head and neck cancer slides. Comparisons are then made for segmentations with and without the proposed EM-based domain adaptation on prostate cancer and head and neck cancer whole slide histopathology images from The Cancer Genome Atlas (TCGA) database and significant improvements are observed.

Deep learning-based predictions of older adults' adherence to cognitive training to support training efficacy.

As the population ages, the number of older adults experiencing mild cognitive impairment (MCI), Alzheimer's disease, and other forms of dementia will increase dramatically over the next few decades. Unfortunately, cognitive changes associated with these conditions threaten independence and quality of life. To address this, researchers have developed promising cognitive training interventions to help prevent or reverse cognitive decline and cognitive impairment. However, the promise of these interventions will not be realized unless older adults regularly engage with them over the long term, and like many health behaviors, adherence to cognitive training interventions can often be poor. To maximize training benefits, it would be useful to be able to predict when adherence lapses for each individual, so that support systems can be personalized to bolster adherence and intervention engagement at optimal time points. The current research uses data from a technology-based cognitive intervention study to recognize patterns in participants' adherence levels and predict their future adherence to the training program. We leveraged the feature learning capabilities of deep neural networks to predict patterns of adherence for a given participant, based on their past behavior. A separate, personalized model was trained for each participant to capture individualistic features of adherence. We posed the adherence prediction as a binary classification problem and exploited multivariate time series analysis using an adaptive window size for model training. Further, data augmentation techniques were used to overcome the challenge of limited training data and enhance the size of the dataset. To the best of our knowledge, this is the first research effort to use advanced machine learning techniques to predict older adults' daily adherence to cognitive training programs. Experimental evaluations corroborated the promise and potential of deep learning models for adherence prediction, which furnished highest mean F-scores of 75.5, 75.5, and 74.6% for the Convolution Neural Network (CNN), Long Short-Term Memory (LSTM) network, and CNN-LSTM models respectively.

SAOCNN: Self-Attention and One-Class Neural Networks for Hyperspectral Anomaly Detection

Hyperspectral anomaly detection is a popular research direction for hyperspectral images; however, it is problematic because it separates the background and anomaly without prior target information. Currently, deep neural networks are used as an extractor to mine intrinsic features in hyperspectral images, which can be fed into separate anomaly detection methods to improve their performances. However, this hybrid approach is suboptimal because the subsequent detector is unable to drive the data representation in hidden layers, which makes it a challenge to maximize the capabilities of deep neural networks when extracting the underlying features customized for anomaly detection. To address this issue, a novel unsupervised, self-attention-based, one-class neural network (SAOCNN) is proposed in this paper. SAOCNN consists of two components: a novel feature extraction network and a one-class SVM (OC-SVM) anomaly detection method, which are interconnected and jointly trained by the OC-SVM-like loss function. The adoption of co-training updates the feature extraction network together with the anomaly detector, thus improving the whole network’s detection performance. Considering that the prominent feature of an anomaly lies in its difference from the background, we designed a deep neural extraction network to learn more comprehensive hyperspectral image features, including spectral, global correlation, and local spatial features. To accomplish this goal, we adopted an adversarial autoencoder to produce the residual image with highlighted anomaly targets and a suppressed background, which is input into an improved non-local module to adaptively select the useful global information in the whole deep feature space. In addition, we incorporated a two-layer convolutional network to obtain local features. SAOCNN maps the original hyperspectral data to a learned feature space with better anomaly separation from the background, making it possible for the hyperplane to separate them. Our experiments on six public hyperspectral datasets demonstrate the state-of-the-art performance and superiority of our proposed SAOCNN when extracting deep potential features, which are more conducive to anomaly detection.

Covid-19 detection using chest X-rays: is lung segmentation important for generalization?

PurposeWe evaluated the generalization capability of deep neural networks (DNNs) in the task of classifying chest X-rays as Covid-19, normal or pneumonia, when trained in a relatively small and mixed datasets.MethodsWe proposed a DNN to perform lung segmentation and classification, stacking a segmentation module (U-Net), an original intermediate module and a classification module (DenseNet201). To evaluate generalization capability, we tested the network with an external dataset (from distinct localities) and used Bayesian inference to estimate the probability distributions of performance metrics. Furthermore, we introduce a novel evaluation technique, which uses layer-wise relevance propagation (LRP) and Brixia scores to compare the DNN grounds for decision with radiologists.ResultsThe proposed DNN achieved 0.917 AUC (area under the ROC curve) on the external test dataset, surpassing a DenseNet without segmentation, which showed 0.906 AUC. Bayesian inference indicated mean accuracy of 76.1% and [0.695, 0.826] 95% HDI (high-density interval, which concentrates 95% of the metric’s probability mass) with segmentation and, without segmentation, 71.7% and [0.646, 0.786].ConclusionEmploying an analysis based on LRP and Brixia scores, we discovered that areas where radiologists found strong Covid-19 symptoms are the most important for the stacked DNN classification. External validation showed smaller accuracies than internal, indicating difficulty in generalization, which is positively affected by lung segmentation. Finally, the performance on the external dataset and the analysis with LRP suggest that DNNs can successfully detect Covid-19 even when trained on small and mixed datasets.

Microseismic Monitoring and Analysis Using Cutting-Edge Technology: A Key Enabler for Reservoir Characterization

Microseismic monitoring is a useful enabler for reservoir characterization without which the information on the effects of reservoir operations such as hydraulic fracturing, enhanced oil recovery, carbon dioxide, or natural gas geological storage would be obscured. This research provides a new breakthrough in the tracking of the reservoir fracture network and characterization by detecting the microseismic events and locating their sources in real-time during reservoir operations. The monitoring was conducted using fiber optic distributed acoustic sensors (DAS) and the data were analyzed by deep learning. The use of DAS for microseismic monitoring is a game changer due to its excellent temporal and spatial resolution as well as cost-effectiveness. The deep learning approach is well-suited to dealing in real-time with the large amounts of data recorded by DAS equipment due to its computational speed. Two convolutional neural network based models were evaluated and the best one was used to detect and locate microseismic events from the DAS recorded field microseismic data from the FORGE project in Milford, United States. The results indicate the capability of deep neural networks to simultaneously detect and locate microseismic events from the raw DAS measurements. The results showed a small percentage error. In addition to the high spatial and temporal resolution, fiber optic cables are durable and can be installed permanently in the field and be used for decades. They are also resistant to high pressure, can withstand considerably high temperature, and therefore can be used even during field operations such as a flooding or hydraulic fracture stimulation. Deep neural networks are very robust; need minimum data pre-processing, can handle large volumes of data, and are able to perform multiple computations in a time- and cost-effective way. Once trained, the network can be easily adopted to new conditions through transfer learning.

Deep MIMO radar target detector in Gaussian clutter

We focus on the target detection problem of the multiple-input multiple-output (MIMO) radar in clutter. Most of the existing MIMO radar detection works rely on the clutter models, which may limit their scopes of application. To address this issue, a deep learning based MIMO radar target detection framework is proposed in this paper, which utilises the powerful representation and discrimination capabilities of deep neural networks (DNNs) to improve the detection performance in a data-driven manner and does not require any prior information about the clutter distribution. In most classification problems, DNNs are trained with the cross entropy loss, which promotes DNNs to achieve higher accuracy. However, the main concern in the radar target detection problem is the probability of detection (PD) under a given probability of false alarm. In this framework, a novel loss is proposed to directly maximise the average PD of the DNN, or equivalently maximise the area under curve of the DNN. Under the DL-based MIMO radar target detection framework, a deep convolutional neural network (CNN) architecture is introduced, and an input feature of the received signal for the CNN based on the conventional generalised likelihood ratio test statistics is designed to further improve the detection performance. Finally, extensive simulation results are presented to validate the detection performance and the robustness of the proposed methods.

Discriminative Visual Similarity Search with Semantically Cycle-consistent Hashing Networks

Deep hashing has great potential in large-scale visual similarity search due to its preferable efficiency in storage and computation. Technically, deep hashing for visual similarity search inherits the powerful representation capability of deep neural networks, and it encodes visual features into compact binary codes by preserving representative semantic visual features. Works in this field mainly focus on building the relationship between the visual and objective hash spaces, while they seldom study the triadic cross-domain semantic knowledge transfer among visual, semantic, and hashing spaces, leading to a serious semantic ignorance problem during space transformation. In this article, we propose a novel deep tripartite semantically interactive hashing framework, dubbed Semantically Cycle-consistent Hashing Networks (SCHNs), for discriminative hash code learning. Particularly, we construct a flexible semantic space and a transitive latent space, in conjunction with the visual space, to jointly deduce the privileged discriminative hash space. Specifically, a new semantic space is conceived to strengthen the flexibility and completeness of categories in the semantic feature inference phase. At the same time, a transitive latent space is formulated to explore and uncover the shared semantic interactivity embedded in visual and semantic features. Moreover, to further ensure semantic consistency across multiple spaces, we propose to build a cyclic adversarial learning module to preserve and keep their semantic concurrence during space transformation. Notably, our SCHN, for the first time, establishes the cyclic principle of deep semantic-preserving hashing by adaptive semantic parsing across different spaces in a single-modal visual similarity search. In addition, the entire learning framework is jointly optimized in an end-to-end manner. Extensive experiments performed on diverse large-scale datasets evidence the superiority of our method against other state-of-the-art deep hashing algorithms. The source codes of this article are available at https://github.com/JalinWang/SCHN.

Deep Learning Techniques to Classify the Aerial Images with Gabor Filter

Abstract: Aerial Images are a valuable data source for earth observation, can help us to measure and observe detailed structures on the Earth’s surface. Aerial images are drastically growing. This has given particular urgency to the quest for how to make full use of ever-increasing Aerial images for intelligent earth observation. Hence, it is extremely important to understand huge and complex Aerial images. Aerial image classification, which aims at labeling Aerial images with a set of semantic categories based on their contents, has broad applications in a range of fields. Propelled by the powerful feature learning capabilities of deep neural networks, Aerial image scene classification driven by deep learning has drawn remarkable attention and achieved significant breakthroughs. However, recent achievements regarding deep learning for scene classification of Aerial images is still lacking. Keywords: Deep learning, Aerial image, CNN, Gabor Filter

User preference mining based on fine-grained sentiment analysis

User preference mining is an application of data mining that attracts increasing attention. Although most of the existing user preference mining methods achieved significant performance improvement, the sentiment tendencies of users were seldom considered. This paper proposes fine-grained sentiment analysis for preference mining. The powerful feature representation capabilities of deep neural networks have significantly improved the performance of fine-grained sentiment analysis. But two main challenges remain when using deep neural network models: incomplete user feature extraction and insufficient interaction. In response, a pre-training language model is employed to encode user features to fully explore potential interests of users, a linguistic knowledge model is introduced to assist the encoding, a multi-scale convolution neural network is adopted to capture text features at different scales and fully utilize the text information, and the fine-grained sentiment analysis task is modeled as a sequence labeling problem to explore the sentiment polarity of user evaluation. Experiments on a user review data set are used to verify the new approach. Experimental results of precision, recall rate and F1-value show that the proposed approach performs better, and is more effective than baseline models. For example, the F1-value is increased by 4.27% compared to the best performing baseline model. Findings have important implications for research and practice.

Tongue Segmentation for Disease Diagnosis a Precise and Fast method using Double U-Net Architecture

Abstract: A robust automatic tongue diagnosis system greatly relies on accurate segmentation of the tongue body from the image. It is a challenge to accurately segment tongue body from an image having close interference of teeth, lips and face. Deep Learning methods such as Deep Convolution Networks like FCN (fully connected Networks), U-NET, Res-Net (Residual Network) have superseded the performance of conventional techniques like Active Contour, Gradient V Flow, Level Set etc. . In this paper looking into the tremendous capability of Deep neural Networks, we have employed Double U-NET for tongue image segmentation of images captured by mobile device and compared the results with that of U-NET, and Res U-NET architectures. Qualitative as well as quantitative analysis of the three reveals a superior performance of Double U-net especially in images with additional dominant features of the face such as lips, teeth, and spaces with the tongue image. Keywords: Double U-Net, Residual U-Net, U-Net, tongue segmentation, automatic tongue diagnosis

MS[formula omitted]GAH: Multi-label semantic supervised graph attention hashing for robust cross-modal retrieval

Due to the strong nonlinear representation capabilities of deep neural networks and the low storage and high efficiency characteristics of hash learning, deep cross-modal hashing has been propelled to the forefront of academics. How to preferably bridge semantic relevance to further bridge the semantic modality gap is the vital bottleneck to improve model performance. Confronting samples with rich semantics, how to comprehensively explore the hidden correlations and establish more precise modality relationships is the primary issue to be solved. In this work, we propose a novel deep hashing method called Multi-Label Semantic Supervised Graph Attention Hashing (MS2GAH), which is an end-to-end framework that integrates graph attention networks (GATs). It constructs graph features through the adjacency of nodes and assigns different weights to adjacent edges to enhance the robustness of the model. Simultaneously, multi-label annotations are utilized to bridge the semantic relevance between modalities in a more fine-grained manner. To make preferable use of rich semantic information, an end-to-end label encoder is designed to mine high-level semantics from multi-label annotations to guide the feature extraction process of specific-modality networks, thereby further narrowing the modality gap. Finally, extensive experiments have been conducted on four datasets, and the results show that MS2GAH is superior to other baselines and one step forward.

Robot navigation in a crowd by integrating deep reinforcement learning and online planning

Navigating mobile robots along time-efficient and collision-free paths in crowds is still an open and challenging problem. The key is to build a profound understanding of the crowd for mobile robots, which is the basis of a proactive and foresighted policy. However, since the interaction mechanisms among pedestrians are complex and sophisticated, it is difficult to describe and model them accurately. For the excellent approximation capability of deep neural networks, deep reinforcement learning is a promising solution to the problem. However, current model-free learning-based approaches in crowd navigation always neglect planning and still lead to reactive collision avoidance policies and shortsighted behaviors. Meanwhile, most model-based learning-based approaches are based on state values, imposing a substantial computational burden. To address these problems, we propose a graph-based deep reinforcement learning method, social graph-based double dueling deep Q-network (SG-D3QN), that (i) introduces a social attention mechanism to extract an efficient graph representation for the crowd-robot state, (ii) extends the previous state value approximator to a state-action value approximator, (iii) further optimizes the state-action value approximator with simulated experiences generated by the learned environment model, and (iv) then proposes a human-like decision-making process by integrating model-free reinforcement learning and online planning. Experimental results indicate that our approach helps the robot understand the crowd and achieves a high success rate of more than 0.99 in the crowd navigation task. Compared with previous state-of-the-art algorithms, our approach achieves better performance. Furthermore, with the human-like decision-making process, our approach incurs less than half of the computational cost.

Estimating the state of epidemics spreading with graph neural networks.

When an epidemic spreads into a population, it is often impractical or impossible to continuously monitor all subjects involved. As an alternative, we propose using algorithmic solutions that can infer the state of the whole population from a limited number of measures. We analyze the capability of deep neural networks to solve this challenging task. We base our proposed architecture on Graph Convolutional Neural Networks. As such, it can reason on the effect of the underlying social network structure, which is recognized as the main component in spreading an epidemic. The proposed architecture can reconstruct the entire state with accuracy above 70%, as proven by two scenarios modeled on the CoVid-19 pandemic. The first is a generic homogeneous population, and the second is a toy model of the Boston metropolitan area. Note that no retraining of the architecture is necessary when changing the model.

Deep plug-and-play prior for hyperspectral image restoration

Deep-learning-based hyperspectral image (HSI) restoration methods have gained great popularity for their remarkable performance but often demand expensive network retraining whenever the specifics of task changes. In this paper, we propose to restore HSIs in a unified approach with an effective plug-and-play method, which can jointly retain the flexibility of optimization-based methods and utilize the powerful representation capability of deep neural networks. Specifically, we first develop a new deep HSI denoiser leveraging gated recurrent convolution units, short- and long-term skip connections, and an augmented noise level map to better exploit the abundant spatio-spectral information within HSIs. It, therefore, leads to the state-of-the-art performance on HSI denoising under both Gaussian and complex noise settings. Then, the proposed denoiser is inserted into the plug-and-play framework as a powerful implicit HSI prior to tackle various HSI restoration tasks. Through extensive experiments on HSI super-resolution, compressed sensing, and inpainting, we demonstrate that our approach often achieves superior performance, which is competitive with or even better than the state-of-the-art on each task, via a single model without any task-specific training.

Graph-based relational reasoning in a latent space for skeleton-based action recognition

Motivated by the powerful capability of deep neural networks in feature learning, a new graph-based neural network is proposed to learn local and global relational information on skeleton sequences represented as spatio-temporal graphs (STGs). The pipeline of our network architecture consists of three main stages. As the first stage, spatial–temporal sub-graphs (sub-STGs) are projected into a latent space in which every point is represented as a linear subspace. The second stage is based on message passing to acquire the localized correlated features of the nodes in the latent space. The third stage relies on graph convolutional networks (GCNs) to reason the long-range spatio-temporal dependencies through a graph representation of the latent space. Finally, the average pooling layer and the softmax classifier are then employed to predict the action categories based on the extracted local and global correlations. We validate our model in terms of action recognition using three challenging datasets: the NTU RGB+D, Kinetics Motion, and SBU Kinect Interaction datasets. The experimental results demonstrate the effectiveness of our approach and show that our proposed model outperforms the state-of-the-art methods.

MS-CNN: multiscale recognition of building rooftops from high spatial resolution remote sensing imagery

ABSTRACT The effective recognition and precise positioning of multiscale building rooftop is one of the key scientific problems that have yet to be resolved urgently in the current implementation of high-resolution remote sensing. In recent years, the automatic recognition of high-resolution image targets often employs convolutional neural networks to extract features. However, such traditional methods often ignore multiscale features of geographical objects, while lacking effective multiscale information extraction strategies. By utilizing the feature learning capability of deep neural networks, this study proposes a multiscale convolutional neural network named MS-CNN to recognize building rooftops from high-resolution remote sensing imagery. In addition, this study constructs a pedigree deep learning sample library based on the remote sensing Tupu theory that considers the spectral and geometric characteristics of building rooftops. Able to utilize feature segmentation mechanism and fusion enhancement strategy, MS-CNN enriches the receptive fields obtained by each convolution layer. The proposed network of this study is also compared with the famous Mask R-CNN method, proving the relative advantages of the MS-CNN method with multiscale characteristics. The experimental results show that the precision and recall metrics of the MS-CNN are 4.18% (.8655 vs. .8238) and 5.71% (.8380 vs. .7809) higher than those of the Mask R-CNN, respectively. The proposed method has been deployed in practical engineering projects in Vietnam and Myanmar, etc.

Panchromatic Side Sparsity Model-Based Deep Unfolding Network for Pansharpening

Deep learning (DL) recently receives state-of-the-art results in pansharpening. However, most of the existing pansharpening neural networks are purely data-driven, without taking account of the characteristics of the pansharpening task. To address this issue, we propose a novel deep unfolding network for pansharpening that combines the insight of the variational optimization (VO) model and the capability of deep neural network (DNN). We first develop a panchromatic side sparsity (PASS) prior-based VO model for pansharpend image reconstruction, which is formulated as the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\ell _{1}-\ell _{1}$ </tex-math></inline-formula> minimization. In particular, the PASS prior is defined using the transform sparsity, which can alleviate the influences of the irregular outliers between the multispectral (MS) and panchromatic (PAN) images. The iterations of half-quadratic splitting algorithm for solving the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\ell _{1}-\ell _{1}$ </tex-math></inline-formula> minimization are then deeply unfolded into a DNN, referred as PASS-Net. To capture the nonlinear relationship between MS and PAN images, the linear transforms used in PASS prior are extended into the subnetworks in PASS-Net. Moreover, a pair of learnable downsampling and upsampling modules are designed to realize the downsampling and upsampling operations, which can improve the flexibility. The experimental results on different satellite datasets confirm that PASS-Net is superior to some representational traditional methods and state-of-the-art DL-based methods.

Identification and Classification of Crowd Activities

The identification and classification of collective people's activities are gaining momentum as significant themes in machine learning, with many potential applications emerging. The need for representation of collective human behavior is especially crucial in applications such as assessing security conditions and preventing crowd congestion. This paper investigates the capability of deep neural network (DNN) algorithms to achieve our carefully engineered pipeline for crowd analysis. It includes three principal stages that cover crowd analysis challenges. First, individual's detection is represented using the You Only Look Once (YOLO) model for human detection and Kalman filter for multiple human tracking; Second, the density map and crowd counting of a certain location are generated using bounding boxes from a human detector; and Finally, in order to classify normal or abnormal crowds, individual activities are identified with pose estimation. The proposed system successfully achieves designing an effective collective representation of the crowd given the individuals in addition to introducing a significant change of crowd in terms of activities change. Experimental results on MOT20 and SDHA datasets demonstrate that the proposed system is robust and efficient. The framework achieves an improved performance of recognition and detection people with a mean average precision of 99.0%, a real-time speed of 0.6 ms non-maximum suppression (NMS) per image for the SDHA dataset, and 95.3% mean average precision for MOT20 with 1.5 ms NMS per image.