Yanxiao Zhao, Yaqian Li, Zi-Hao Bo, Rinyoichi Takezoe, Haojia Hui, Mo Guang, Renlei, Xiaolin Qin, Kaiwen Long. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

We tackle monocular 6D pose estimation for small and weakly textured objects using a Transformer enhanced with geometry-aware priors. Low-Rank Gated Module applies a low rank bottleneck with a learned gate to denoise query features and to amplify fine detail. Topology Optimization builds a Gaussian affinity graph over queries, pools a global summary, and broadcasts it back while enforcing a top‑k connectivity loss that promotes coherent clusters. A training-time Riemannian Refinement performs a few geodesic steps on SO(3) with learned step sizes and per query weights, keeping rotations on the manifold and stabilizing translation. Our methods achieved excellent results on three publicly available datasets, and we propose a new small objects 6DoF estimation dataset for UAVs.

Aspect-based sentiment analysis (ABSA) aims to extract fine-grained opinions from the text by discerning sentiments toward specific aspects. Although large language models (LLMs) perform well in-context learning (ICL), current ICL methodologies typically retrieve semantically similar but structurally redundant examples, failing to capture syntactic and aspect-level cues critical for ABSA. To overcome these limitations, we report Multi-perspective Sequential retrieval with Predictive Feedback (MSPF), a few-shot learning framework that enhances ICL through MSPF, which integrates three complementary perspectives: overall semantic, syntactic relevance, and aspect sentiment alignment. Evaluated on four benchmark datasets (Laptop14, Restaurant14, Books, and Clothing), MSPF achieved F1 scores of 67.03 % (Laptop14), 73.51 % (Restaurant14), 76.07 % (Books), and 81.96 % (Clothing), outperforming standard ICL by +7.06 %, +5.60 %, +25.61 %, and +18.38 %, respectively. These results validated the efficacy of MSPF in improving LLM reasoning for fine-grained sentiment tasks with limited annotations.

Low Earth Orbit (LEO) satellite constellations are crucial for enabling global coverage and low-latency services. Satellite Edge Computing enables on-board processing to improve task responsiveness and reduce backhaul load. Microservices, with their modular and lightweight design, naturally fit the dynamic and constrained satellite environment. However, deploying microservices on LEO satellites faces unique challenges, including stringent energy constraints, highly dynamic connectivity, and complex service dependencies across mobile nodes. In this paper, we propose a Microservice Deployment Framework for Satellite Edge Computing (SEC-MDF). We construct a microservice deployment model that includes real track dynamics, service dependencies, resource constraints, and data transfer patterns, and design a deployment strategy based on deep reinforcement learning (DRL). This strategy is augmented by Heuristic-based Episodic Reward Optimization (HERO), a tailored reward optimization mechanism for SEC-MDF. By integrating A*-based heuristic cost estimation, stage-aware episodic reward buffer, and adaptive reward normalization, HERO significantly enhances the performance of the deployment. Extensive experiments against both heuristic algorithms and DRL variants demonstrate that our framework reduces end-to-end latency by 40 % and system energy consumption by 29 %, significantly outperforming existing approaches. Ablation studies further confirm the critical contribution of the HERO mechanism to overall performance.

Rapid data growth and frequent updates challenge storage system reliability and scalability. This paper introduces Binary Random Generated Codes (BRGCs), highly scalable horizontal codes generated automatically via algebraic computation. BRGCs address limited scalability and high complexity in traditional array codes by constructing optimally sparse generator and parity-check matrices using low-Hamming-weight$\mathrm{m}$-order invertible matrices over the binary field, overcoming prime-number constraints in conventional RAID6 codes. Compared to RDP, EVENODD, Liberation, N, and EAR codes, BRGCs excel in scalability, encoding, decoding, and update complexity. Experimental results indicate that as stripe depth m varies, BRGCs enhance encoding performance$\text{2 3. 3 8 \%}, \text{3 4. 5 2 \%}, \text{2 5. 8 2 \%}, \text{2 1. 9 2 \%}$and 23.22 %, and decoding performance by$16.38 \%, 28.18 \%, 33.57 \%, 16.17 \%$and 19.38 %, relative to the RDP, EVENODD, Liberation, N, and EAR codes, respectively. With a fixed stripe depth m, BRGCs improve encoding and decoding performance by 4.69 % to 43.47 % and 4.16 % to 46.84 % over the RDP, EVENODD, N, and EAR codes, and boosts decoding performance by approximately 8.59 % compared to the Liberation code. As the number of disks increases, the write overhead of BRGCs closely matches that of Liberation, N, and EAR codes, while being 31.4% and 30.7% lower than RDP and EVENODD codes, respectively.

With the rapid development of autonomous driving technology, automated valet parking has imposed more stringent requirements on environmental perception, especially in the demanding underground parking scenarios featuring complex structures and diverse lighting conditions. Semantic segmentation, as a pivotal part of environmental perception, enables the vehicle to better understand its surroundings by assigning semantic labels to every pixel in the image. This study delves into the application of multi-scale augmentation techniques in semantic segmentation, pinpoints their limitations in underground parking scenarios, and puts forward two novel strategies: post-maximum augmentation and post-minimum augmentation. Experiments carried out on the AVM-SemSeg dataset validate that these strategies remarkably boost the model’s performance in such scenarios. Based on this, the post-min-max (PMM) augmentation method has been further developed, integrating the advantages of the two aforementioned novel strategies. It elevates the model’s predictive confidence in true positive samples and curbs the incidence of false positive misclassifications as background. The computational burden is also alleviated by reducing the scale inputs, which enhances the efficiency of the model’s multi-scale inference while still maintaining competitive results. The results manifest that the PMM method outperforms existing multi-scale inference information fusion methods in semantic segmentation tasks and can be combined with all the tested methods to notably enhance their performance on this dataset. This research not only augments the semantic segmentation capabilities of autonomous vehicles in underground parking environments but also offers vital technical support for automated valet parking tasks.

3D Gaussian Splatting (3DGS) has been recognized as a pioneering technique in scene reconstruction and novel view synthesis. Recent work on reconstructing the 3D human body using 3DGS attempts to leverage prior information on human pose to enhance rendering quality and improve training speed. However, it struggles to effectively fit dynamic surface planes due to multi-view inconsistency and redundant Gaussians. This inconsistency arises because Gaussian ellipsoids cannot accurately represent the surfaces of dynamic objects, which hinders the rapid reconstruction of the dynamic human body. Meanwhile, the prevalence of redundant Gaussians means that the training time of these works is still not ideal for quickly fitting a dynamic human body. To address these, we propose EfficientHuman, a model that quickly accomplishes the dynamic reconstruction of the human body using Articulated 2D Gaussian while ensuring high rendering quality. The key innovation involves encoding Gaussian splats as Articulated 2D Gaussian surfels in canonical space and then transforming them to pose space via Linear Blend Skinning (LBS) to achieve efficient pose transformations. Unlike 3D Gaussians, Articulated 2D Gaussian surfels can quickly conform to the dynamic human body while ensuring view-consistent geometries. Additionally, we introduce a pose calibration module and an LBS optimization module to achieve precise fitting of dynamic human poses, enhancing the model's performance. Extensive experiments on the ZJU-MoCap dataset demonstrate that EfficientHuman achieves rapid 3D dynamic human reconstruction in less than a minute on average, which is ≈36% faster than the current state-of-the-art, while also reducing the number of redundant Gaussians.

Visible-infrared object detection has gained significant attention because of its applications in autonomous driving, video surveillance, and related fields. The effective fusion of multimodal information is fundamental to its success. The existing approaches concentrate on improving the pixel-level fusion mechanisms; detection performance has reached a plateau. We propose a new framework for SAM-guided semantic knowledge fusion ( SemFusion ). The core idea is to leverage semantic priors from large models while incorporating a lightweight cross-modal fusion strategy. Specifically, our method comprises two stages. In the first stage, the Flow-Guided RGB Feature Alignment (FGRA) module establishes object-aware correspondences between multimodalities based on SAM-generated masks. This ensures semantic-level feature matching by deformable convolution alignment. In the second stage, the Semantic Knowledge Distillation (SKD) strategy facilitates the transfer of large-model knowledge to the detection model through SAM feature, offset, and mask level distillations. For the detector model, three blocks are designed to augment any off-the-shelf detector. They are deformable cross-modal alignment, spatio-channel preliminary fusion, and mask-guided feature refinement. By alignment with SAM masks, semantic alignment and fusion can be achieved, breaking the pixel-level fusion barrier. Extensive experiments demonstrate that our method, as a plugin, exhibits superior performance on the DroneVehicle, VEDAI, and LLVIP datasets. Code is available at https://github.com/liting1018/SemFusion.

Cross machine traffic caused by distributed microservice deployment in edge computing significantly affects service performance. And the dynamics of the edge computing, such as fluctuating user request patterns and different network delay make static scheduling stategies challenging. To address these two issue, we first propose a Cross-Machine Traffic-Aware Scheduling Algorithm (CTSA), which models the microservice deployment process as a Markov Decision Process and utilizes a Dueling DQN-based approach to minimize cross-machine traffic while balancing node resource usage. Furthermore, we propose a Dynamic QoS-Aware Scheduling Framework (DQSF) that adapts deployment decisions in real time based on system monitoring to address the challenge of dynamics. Experimental evaluations using a real-world microservice application show that our approach significantly reduces request response time up to 30.2%, improves throughput to 36.7% and ensure Quality of Service (QoS) under dynamic edge computing continuum.

In multi-user edge intelligence system, achieving efficient and stable DNN inference under heterogeneous task priorities is a critical challenge. This paper presents a priority-aware edge-device collaborative inference scheme that models the entire inference workflow while explicitly incorporating user-level priority constraints. The optimization objective is twofold: to maximize, in priority order, the number of users with stable local queue under the strict constraint of server queue stability, and to minimize the total end-to-end latency across all users. A key component is the Server Queuing Latency Estimation (SQLE) algorithm, which decomposes the latency contributions of user priority interactions and iteratively estimates task queuing times. Compared with classical models such as M/D/1, SQLE achieves significantly higher accuracy under dynamic workloads. Based on the predicted latency, we further develop a two-stage offloading decision algorithm: Maximum User Prioritized Selection for Local Queue Stability (MUPS) determines the maximal subset of users whose local queues can be stabilized, and Fine-grained partition point for Latency Optimization (FOPL) refines offloading points to minimize global latency. Experiments on a heterogeneous edge-device system show that SQLE consistently achieves over 90% estimation accuracy with low error variance across varying system scales, significantly outperforming classical queuing models. Under different load, MUPS and FOPL supports more stable users and reduces average end-to-end latency, demonstrating its robustness and superiority over state-of-the-art methods.

Unmanned aerial vehicles (UAVs) have attracted widespread attention in post-disaster search and rescue (SAR) due to high flexibility and low-cost advantages. However, traditional centralized control approaches face problems such as poor adaptability and low robustness in complex and dynamic post-disaster environments. In order to improve the autonomous and execution efficiency of UAV cooperative search tasks, decentralized control methods have gradually become a research focus. However, how to efficiently realize autonomous path planning for UAVs under the condition of limited computational resources is still a key challenge to be solved. In this paper, we propose a dynamic adaptive path optimization method based on evolutionary neural network (DAPO-ENN), which combines the global search capability of evolutionary algorithms with the adaptive characteristics of neural networks to realize the centerless autonomous path planning and search coverage optimization of UAVs in post-disaster environments. DAPO-ENN can optimize the performance of the model under the limitation of computational resources, and adapt to the dynamic changes of the environment by online path optimization adjustment, so as to effectively improve the coverage efficiency while ensuring a high search coverage rate. The experimental results show that the DAPO-ENN proposed in this paper has stronger environmental adaptability and lower resource consumption than the existing comparison algorithms. The results suggest that the method provides an efficient and flexible solution for the cooperative search of UAVs after disasters.

We investigate the problem of path planning and task offloading for UAV clusters in a UAV-assisted edge computing scenario. UAVs autonomously make decisions regarding path planning, continuous service provision, and task offloading based on collected information. In this setting, terminal equipment (TE) cannot directly connect to servers; thus, UAVs act as both edge servers and communication relays, proactively providing services to TEs. We construct a fine-grained temporal scale model that decomposes UAV actions into atomic time units, transforming decision-making on specific behaviors into state transition decisions. This approach better accommodates the needs of time-varying environments. Regarding path planning, given the time-sensitive nature of TE requests, we focus on how to provide stable and timely computational services to TEs. We propose a multi-agent reinforcement learning algorithm capable of dynamically sensing task priorities to enhance Quality of Service (QoS), with optimizations made in terms of task completion rate, UAV energy consumption, and processing delay. About task offloading decision, we introduce a dual-keyword-based offloading algorithm to optimize the binary offloading process. Finally, we conduct simulation experiments to demonstrate the effectiveness of the proposed algorithms, and comparative experiments confirm their superiority.

Abstract Estimating the probability of events is a significant challenge in many fields, often requiring a probabilistic model or additional labels and tasks for accurate prediction. However, those methods have limited scalability or unnecessary computational resource consumption due to predicting unrelated values. To address these issues, we propose a novel approach that estimates event probabilities based on the distributions of their first occurrence in the time domain. By using Signal Temporal Logic formulas to describe events and applying an algorithm that estimates complex events’ probabilities through simple event occurrence distributions, this study presents an efficient approach that does not depend on high-precision prediction. We evaluate the performance of our method on simulated scenarios of unmanned aerial vehicle motion and autonomous driving.

Glass largely blurs the boundary between the real world and the reflection. The special transmittance and reflectance quality have confused the semantic tasks related to machine vision. Therefore, how to clear the boundary built by glass, and avoid over-capturing features as false positive information in deep structure, matters for constraining the segmentation of reflection surface and penetrating glass. We propose the Fourier Boundary Features Network with Wider Catchers (FBWC), which might represent the first attempt to utilize sufficiently wide horizontal shallow branches without vertical deepening for guiding the fine granularity segmentation boundary through primary glass semantic information. Specifically, we design the Wider Coarse-Catchers (WCC) for anchoring large area segmentation and reducing excessive extraction from a structural perspective. We embed fine-grained features by Cross Transpose Attention (CTA), which is introduced to avoid the incomplete area within the boundary caused by reflection noise. For excavating glass features and balancing high-low layers context, a learnable Fourier Convolution Controller (FCC) is proposed to regulate information integration robustly. The proposed method is validated on three different public glass segmentation datasets. Experimental results reveal that the proposed method yields better segmentation performance compared with the state-of-the-art (SOTA) methods in glass image segmentation.

Deep Neural Network (DNN) has been widely used in intelligence applications due to its excellent performance in executing inference tasks. Since DNN tasks require a large amount of computation and high-resolution raw input, collaborative inference is proposed to partition the DNN model from the middle layer to minimize the end-to-end latency. However, when existing work attempts to combine collaborative inference with task offloading in cloud-edge-end systems, two challenges result in poor latency and throughput: excessive layers in the model make it difficult to find suitable partitions, and multiple decision variables make the reinforcement learning agent challenging to converge. This paper aims to reduce the long-term average end-to-end latency of DNN tasks by jointly optimizing task offloading, model partitioning, and resource allocation in dynamic environments. To solve the problem of failing to find suitable partitions, we propose a novel Optional Partition Point Compression (OPPC) algorithm, which selects the high-quality partition points based on layers’ output feature to reduce the difficulty of model partitioning. To improve the convergence performance of the agent, we propose a Reinforcement Learning based Collaborative Inference Optimization (RLCIO) algorithm. Unlike the existing reinforcement learning architecture, RLCIO decouples resource allocation using an Edge Computing Resource Allocation (ECRA) algorithm to reduce the agent’s decision variables. Simulation results show that the RLCIO algorithm performs better than five related schemes, reduces the average end-to-end latency of system tasks by 72% and improves the system throughput by 3.5x in the best case.

The prediction of events in a specific time interval is of great significance in different fields such as finance, health care and disaster early warning. In addition to common end-to-end models, researchers have explored event sequence prediction (a type of multi-label classification models) and time series fragment forecasting (a type of sequence forecasting models) within prediction intervals. However, these approaches often accumulate errors when dealing with larger intervals or require retraining to adapt to the same event occurring within a different interval. To overcome these limitations, we introduce the time to event models into event prediction, and utilize the distributions of events’ first occurrences within a broader time domain for making predictions. By focusing on this extended temporal context, our method aims to provide improved predictions for events across different intervals. We evaluate the performance of our proposed approach on a simulated dataset and three realistic datasets representing distinct domains. Our results show that our method performs comparable to end-to-end methods, and due to the nature of the probability density function, it does not necessitate retraining when adjusting interval lengths.

Accurate localization in challenging garage environments—marked by poor lighting, sparse textures, repetitive structures, dynamic scenes, and the absence of GPS—is crucial for automated valet parking (AVP) tasks. Addressing these challenges, our research introduces AVM-SLAM, a cutting-edge semantic visual SLAM architecture with multi-sensor fusion in a bird’s eye view (BEV). This novel framework synergizes the capabilities of four fisheye cameras, wheel encoders, and an inertial measurement unit (IMU) to construct a robust SLAM system. Unique to our approach is the implementation of a flare removal technique within the BEV imagery, significantly enhancing road marking detection and semantic feature extraction by convolutional neural networks for superior mapping and localization. Our work also pioneers a semantic prequalification (SPQ) module, designed to adeptly handle the challenges posed by environments with repetitive textures, thereby enhancing loop detection and system robustness. To demonstrate the effectiveness and resilience of AVM-SLAM, we have released a specialized multi-sensor and high-resolution dataset of an underground garage, accessible at https://yale-cv.github.io/avm-slamdataset, encouraging further exploration and validation of our approach within similar settings.

There are two mainstream strategies for image-text matching at present. The one, termed as joint embedding learning, aims to model the semantic information of both image and sentence in a shared feature subspace, which facilitates the measurement of semantic similarity but only focuses on global alignment relationship. To explore the local semantic relationship more fully, the other one, termed as metric learning, aims to learn a complex similarity function to directly output score of each image-text pair. However, it significantly suffers from more computation burden at retrieval stage. In this paper, we propose a hierarchically joint embedding model to incorporate the local semantic relationship into a joint embedding learning framework. The proposed method learns the shared local and global embedding spaces simultaneously, and models the joint local embedding space with respect to specific local similarity labels which are easy to access from the lexical information of corpus. Unlike the methods based on metric learning, we can prepare the fixed representations of both images and sentences by concatenating the normalized local and global representations, which makes it feasible to perform the efficient retrieval. And experiments show that the proposed model can achieve competitive performance when compared to the existing joint embedding learning models on two publicly available datasets Flickr30k and MS-COCO.

There is a great need to accurately predict short-term precipitation, which has socioeconomic effects such as agriculture and disaster prevention. Recently, the forecasting models have used multisource data as the multimodality input, thus improving the prediction accuracy. However, the prevailing methods usually suffer from the desynchronization of multisource variables, the insufficient capability of capturing spatiotemporal dependency, and unsatisfactory performance in predicting extreme precipitation events. To fix these problems, we propose a short-term precipitation forecasting model based on spatiotemporal alignment attention, with self-attention for temporal alignment (SATA) as the temporal alignment module and spatiotemporal attention unit (STAU) as the spatiotemporal feature extractor to filter high-pass features from precipitation signals and capture multiterm temporal dependencies. Based on satellite and ERA5 data from the southwestern region of China, our model achieves improvements of 12.61% in terms of root mean square error (RMSE), in comparison to the state-of-the-art methods.

In an intelligent transportation system, accurate traffic flow prediction can provide significant help for travel planning. Even though some methods are proposed to do so, they focus on either algorithm or data level studies. This work focuses on both by proposing a Community-based dandelion algorithm-enabled Feature selection and Broad learning system (CFB). Specifically, a feature selection method is adopted to choose suitable features aiming to avoid redundant ones affecting prediction accuracy, and a neural network-based learning algorithm, namely a Broad Learning System (BLS), is used to predict traffic flow. In order to further boost its prediction performance, a Community-based Dandelion Algorithm (CDA) is proposed by considering an individual and its multiple offspring as a community and adopting a learning strategy for different communities. The proposed CDA is used to a) choose the suitable features as a feature selection method; and b) optimize the parameters and network structure of BLS. CDA’s superiority over its competitive peers is first verified on CEC2013’s benchmark functions, and then the proposed CFB is applied to handle the traffic flow prediction problems. The results indicate that it can improve the prediction accuracy by 5%-16% compared to the updated traffic flow prediction methods.

To satisfy the differentiated service demands of delay-sensitive and computation-intensive tasks, there is an urgent need to efficiently allocate limited network resources to improve content distribution in cloud-edge collaboration environments. Task partition, as a novel processing scheme, has emerged in people’s vision, but there is little discussion on the application of task partitioning in cloud edge collaborative environments in existing work. In this paper, we propose a dynamic caching scheme based on task partition to optimize resource allocation in a cache-assisted cloud-edge collaboration environment. Specifically, we formulate the task-partition-based optimal content caching problem as a delay minimization model in cloud-edge collaboration system. A low-complexity algorithm based on Dual Decomposition (DD) is proposed to make optimal content caching decisions based on the current network state. Simulation results show that, compared with existing cloudside cooperative caching network models, this model shows advantages in terms of low delay and low complexity.

Loosely coupled and highly cohesived microservices running in containers are becoming the new paradigm for application development. Compared with monolithic applications, applications built on microservices architecture can be deployed and scaled independently, which promises to simplify software development and operation. However, the dramatic increase in the scale of microservices and east-west network traffic in the data center have made the cluster management more complex. Not only does the scale of microservices cause a great deal of pressure on cluster management, but also cascading QoS violations present a substantial risk for SLOs (Service Level Objectives). In this paper, we propose a Microservice-Oriented Topology-Aware Scheduling Framework (MOTAS), which effectively utilizes the topologies of microservices and clusters to optimize the network overhead of microservice applications through a heuristic graph mapping algorithm. The proposed framework can also guarantee the cluster resource utilization. To deal with the dynamic environment of microservice, we propose a mechanism based on distributed trace analysis to detect and handle QoS violations in microservice applications. Through real-world experiments, the framework has been proved to be effective in ensuring cluster resource utilization, reducing application end-to-end latency, improving throughput, and handling QoS violations.

Service mesh is a promising micro-services architecture due to its excellent governance capabilities. Unlike traditional service invocation, configurations for governance need to be issued in the service mesh. However, we find that the control-plane traffic of governance is distributed in full by default, i.e., each service in the data plane receives all configurations. The vast majority of the configurations are redundant for a specific service. Hence, it is important and challenging to make the control plane aware of the calling relationships between services. In this paper, we propose a traffic management mechanism named DATM. Using this mechanism, the entire cluster can be dynamically controlled and services can be configured on demand. It is implemented through a dependency-aware controller and monitors. The controller first processes the information listened to by the monitors and then analyzes the connection between the metrics and the service requests through intelligent algorithms. Finally, the control traffic for regulating the control plane is generated. Our proposed mechanism is experimentally compared with the default strategy and existing work across a wide set of load scenarios in a testbed based on Istio service mesh and Kubernetes. Experimental results demonstrate that our mechanism can save the storage resources of a single agent by 40% to 60%, and the number of cluster updates can be greatly reduced. From the perspective of the whole cluster, the optimization results are even better.

Service Mesh has become one of the most popular microservices governance frameworks. In the service governance function of traffic management, how to design an efficient strategy for request distribution to minimize response time has attracted wide research interest. Generally, a traffic policy can be pre-set manually. However, such a static policy always performs terribly when the load changes dynamically. That is, Service Mesh cannot adaptively select the optimal policy when setting service version diversion and selecting instances on the service subset. To address this issue, in this paper, we designed an adaptive strategy based on split-flow to achieve global load balancing. Furthermore, we implement such a strategy by developing a plug-in named DTMA for Istio, a typical representative of Service Mesh. By obtaining the Service Mesh node’s topology and global instance load distribution in real-time, Istio updates the optimal split-flow weight and selects the suitable load balancing strategy. Extensive experimental results show that our algorithm can reduce the response time by nearly half and keep a load of nodes and instances stable compared with Istio’s native configuration. The optimization effect of minimizing the response time is achieved through real-time load balancing.

Edge computing provides an opportunity to improve the quality of service (QoS) of Artificial Intelligence (AI) apps for the Internet of Things (IoTs) scenarios. It is an important way to improve the QoS of intelligent apps by deploying Deep Neural Network (DNN) models on edge nodes. Though the DNN execution time affects the QoS of apps significantly. Due to the limited and dynamic edge resources, and sudden load to edge nodes, it is hard to guarantee the DNN execution efficiency. In this paper, we conduct fine-grained decomposition of DNN tasks and propose a Cloud Edge Collaborative Dynamic Task Scheduling mechanism based on DNN layer-partitioning technique. The approach can realize the collaborative computing of DNN models between cloud and edge, and improve the execution efficiency of DNN models, which guarantees the QoS of AI apps. Through simulation experiments, compared with the existing task scheduling mechanism and AI app deployment mode, we show that the proposed cloud edge collaborative dynamic task scheduling mechanism can effectively reduce the average service response time in the edge intelligent system, so as to improve the apps' overall QoS of the system. Meanwhile, the task scheduling mechanism designed in this paper makes it possible for more complex intelligent models to run in a resource-constrained edge environment.

Abstract As one of the prevalent components, feature pyramid network (FPN) is widely used in current object detection models for improving multi‐scale object detection performance. However, its feature fusion mode is still in a misaligned and local manner, thus limiting the representation power. To address the inherited defects of FPN, a novel architecture termed content‐augmented feature pyramid network (CA‐FPN) is proposed in this paper. Firstly, a global content extraction module (GCEM) is proposed to extract multi‐scale context information. Secondly, lightweight linear spatial Transformer connections are added in the top‐down pathway to augment each feature map with multi‐scale features, where a linearized approximate self‐attention function is designed for reducing model complexity. By means of the self‐attention mechanism in Transformer, it is no longer needed to align feature maps during feature fusion, thus solving the misaligned defect. By setting the query scope to the entire feature map, the local defect can also be solved. Extensive experiments on COCO and PASCAL VOC datasets demonstrated that the CA‐FPN outperforms other FPN‐based detectors without bells and whistles and is robust in different settings.

The Electronic Health Record (EHR) has been widely used in cloud-based medical data platforms. Since the owner of the EHR is a patient and the manager is a doctor (or hospital), this separation of ownership of privacy data leads to privacy leakages of the EHR stored in the cloud environment. The tamper-proof and traceable features of the blockchain make it a promising way to solve EHR privacy protection issues. Nevertheless, the latest research findings on blockchain-based schemes for EHR privacy preservation are focused on single blockchain that corresponds to a designated medical institution, which are not compatible with the privacy anti-leakage demands since private data is transferred through multiple blockchains. In this regard, we put forward relay-chain as a service (RaaS), and propose a cross-blockchain based EHR Privacy-preserving scheme (CEPS), which uses relay-chain to achieve secure access to EHR data when patients visit different hospitals. Furthermore, our scheme ensures that patients can delete the link of EHR freely and effectively. Security analysis and performance evaluations are performed, which prove that CEPS is highly secure and efficient impressively.

The Non-Volatile Memory (NVM) is the promising device to store data and accelerate big data analysis due to its excellent I/O performance. However, we find that simply replacing Hard Disk Drive (HDD) with NVM cannot bring the expected performance improvement. In this paper, we take the data recovery issue in Hadoop File System (HDFS) as a case study to investigate how to take advantage of the performance of NVM. We analyze the data recovery mechanism in HDFS and find that the configuration of replication tasks in the DataNode can affect the data recovery significantly. We conduct extensive analysis and experiments to tuning the configuration and also get some interesting findings. With the new configuration, we increase the data recovery performance improvement from 17% to 71%. At the same time, we can also improve the execution performance of MapReduce tasks to 28% to 59% through optimized configuration.

Real-time surgery workflow automatic detection as computer-assisted surgery systems has become an emerging trend due to improving patient safety during surgery. Currently, the convolutional neural networks can show the best performance for content-based video analysis of surgical workflow. In this paper, a novel solution of surgery workflow detection during the procedure was presented, the edge information of original phases from video frames was extracted and then employed to train together with original phases by using a ResNet. Finally, the methods were evaluated on cataract-101 dataset, a publicly available dataset for surgical phase analysis, on which a maximum accuracy of 90.1% was reached. Additionally, the accuracy of 3% improvement was achieved when compared with the method of no processing the data by edge detection. It is shown that using the edge information of original images could improve the performance of surgical phase recognition, because it can be complementary information for original images to recognize the surgical workflow. This paper shows valuable potential to develop modern medical diagnosis and treatment in automating workflow recognition, and the edge processing of original phases for recognition images can also produce new features to assist the network to recognize the original images. Furthermore, the technology studied in this paper can also be used in other video analysis tasks, or classification of image tasks.

The emerging Non-Volatile Memory (NVM) technology has given rise to an opportunity to accelerate big data analysis. In this paper, we investigate the joint job and data scheduling problem in private cloud data center with a hybrid storage system, and we propose Labeling Scheduler, a flexible labeling-based approach for jointly scheduling. The core idea of the approach is to introduce the labeling system to characterize the features of big data analysis jobs and data objects, and conduct data replacement dynamically between NVM and disk. To the best of our knowledge, this is the first work to introduce the labeling methodology to the big data analysis problem in the cloud data center with a hybrid storage system. We conduct extensive simulations and the simulation results show that the Labeling Scheduler has a significant improvement on system utility compared to the method without labeling information. In addition, the Labeling Scheduler guarantees a high NVM hit rate, which is valuable for NVM endurance enhancement.

Attribute reduction is an important process in many fields, such as knowledge discovery, data mining, machine learning, pattern recognition and so on. However, totally labeled data are quite hard to obtain in real life. Thus, we often have to face the situation that not all the data have been associated with labels in advance. Due to the coexistence of labeled and unlabeled data, attribute reduction problem for partially labeled data becomes more complex and challenging. Many scholars have devoted themselves to solve this problem in the past few years. But current algorithms for partially labeled data are not efficient enough in terms of time complexity. To address this issue, we propose a hypergraph model, where two types of induced hypergraphs are designed from partially labeled decision systems. Then, a fast algorithm based on low-complexity heuristics is raised to compute the minimum vertex cover of a hypergraph. Finally, we propose two types of hypergraph models-based attribute reduction algorithms for partially labeled decision systems. Experimental results on broadly used data sets show that the feasibility and efficiency of our proposed algorithms.

In this paper, we would like to demonstrate an intelligent traffic analytics system called T4, which enables intelligent analytics over real-time and historical trajectories from vehicles. At the front end, we visualize the current traffic flow and result trajectories of different types of queries, as well as the histograms of traffic flow and traffic lights. At the back end, T4 is able to support multiple types of common queries over trajectories, with compact storage, efficient index and fast pruning algorithms. The output of those queries can be used for further monitoring and analytics purposes. Moreover, we train the deep models for traffic flow prediction and traffic light control to reduce traffic congestion. A preliminary version of T4 is available at https://sites.google.com/site/shengwangcs/torch.

In this paper, we study the problem of large-scale trajectory data clustering,k-paths, which aims to efficiently identifyk"representative" paths in a road network. Unlike traditional clustering approaches that require multiple data-dependent hyperparameters,k-paths can be used for visual exploration in applications such as traffic monitoring, public transit planning, and site selection. By combining map matching with an efficient intermediate representation of trajectories and a noveledge-based distance(EBD) measure, we present a scalable clustering method to solvek-paths. Experiments verify that we can cluster millions of taxi trajectories in less than one minute, achieving improvements of up to two orders of magnitude over state-of-the-art solutions that solve similar trajectory clustering problems.

This paper presents a new trajectory search engine called Torch for querying road network trajectory data. Torch is able to efficiently process two types of typical queries (similarity search and Boolean search), and support a wide variety of trajectory similarity functions. Additionally, we propose a new similarity function LORS in Torch to measure the similarity in a more effective and efficient manner. Indexing and search in Torch works as follows. First, each raw vehicle trajectory is transformed to a set of road segments (edges) and a set of crossings (vertices) on the road network. Then a lightweight edge and vertex index called LEVI is built. Given a query, a filtering framework over LEVI is used to dynamically prune the trajectory search space based on the similarity measure imposed. Finally, the result set (ranked or Boolean) is returned. Extensive experiments on real trajectory datasets verify the effectiveness and efficiency of Torch.

In this paper, we build a trip planning system called TISP, which enables user's interactive exploration of POIs and trajectories in their incremental trip planning. At the back end, TISP is able to support seven types of common queries over spatial-only, spatial-textual and textual-only data, based on our proposed unified indexing and search paradigm [7]. At the front end, we propose novel visualisation designs to present the result of different types of queries; our user-friendly interaction designs allow users to construct further queries without inputting any text.

We study a new type of spatial-textual trajectory search: the Exemplar Trajectory Query (ETQ), which specifies one or more places to visit, and descriptions of activities at each place. Our goal is to efficiently find the top-k trajectories by computing spatial and textual similarity at each point. The computational cost for pointwise matching is significantly higher than previous approaches. Therefore, we introduce an incremental pruning baseline and explore how to adaptively tune our approach, introducing a gap-based optimization and a novel twolevel threshold algorithm to improve efficiency. Our proposed methods support order-sensitive ETQ with a minor extension. Experiments on two datasets verify the efficiency and scalability of our proposed solution.

Detecting outlier in spatial database is important for many KDD applications. Existing works in outlier detection don't distinguish between spatial dimension and non-spatial dimension or have poor efficiency. In this paper, we proposed a new measure to identify spatial outliers. We defined spatial outlier factor (SOF) to detect spatial outliers efficiently, and proposed a algorithm (SOFind) to identify them. SOF can successfully identify significant outliers and filtrate some meaningless outliers but can't do it by other methods. The experimental results show that our approach is effective and efficient.