- Topic Modeling
- Target Tracking and Data Fusion in Sensor Networks
- Higher Education and Teaching Methods
- Natural Language Processing Techniques
- Ideological and Political Education
- Innovative Educational Techniques
- Human Pose and Action Recognition
- Multimodal Machine Learning Applications
- Underwater Vehicles and Communication Systems
- Video Surveillance and Tracking Methods
- Gait Recognition and Analysis
- Robotic Path Planning Algorithms
- Advanced Vision and Imaging
- Advanced Text Analysis Techniques
- Advanced Wireless Network Optimization
- Educational Technology and Assessment
- Chaos control and synchronization
- Guidance and Control Systems
- Educational Technology and Pedagogy
- Neural Networks Stability and Synchronization
- Education and Work Dynamics
- Handwritten Text Recognition Techniques
- Engineering Education and Curriculum Development
- IoT-based Smart Home Systems
- Military Defense Systems Analysis
Zhejiang University
2013-2024
Carnegie Mellon University
2020-2023
Tsinghua University
2018-2021
University Town of Shenzhen
2019-2021
Nanyang Medical College
2008-2019
East China University of Science and Technology
2013-2018
University of Electronic Science and Technology of China
2016
University of Chinese Academy of Sciences
2013
Tongji University
2012
China University of Geosciences (Beijing)
2011
This paper presents CLEAR, a retrieval model that seeks to complement classical lexical exact-match models such as BM25 with semantic matching signals from neural embedding model. CLEAR explicitly trains the encode language structures and semantics fails capture novel residual-based learning method. Empirical evaluations demonstrate advantages of over state-of-the-art models, it can substantially improve end-to-end accuracy efficiency reranking pipelines.
In this paper, H∞ synchronization and state estimation problems are considered for different types of chaotic systems. A unified model consisting a linear dynamic system bounded static nonlinear operator is employed to describe these systems, such as Hopfield neural networks, cellular Chua's circuits, Qi recurrent multilayer perceptrons, etc. Based on the performance analysis using matrix inequality approach, novel feedback controllers established not only guarantee exponentially stable...
Egocentric human pose estimation (HPE) using wearable sensors is essential for VR/AR applications. Most methods rely solely on either egocentric-view images or sparse Inertial Measurement Unit (IMU) signals, leading to inaccuracies due self-occlusion in the sparseness and drift of inertial sensors. importantly, lack real-world datasets containing both modalities a major obstacle progress this field. To overcome barrier, we propose EMHI, multimodal Motion dataset with Head-Mounted Display...
As an important task in multimodal context understanding, Text-VQA (Visual Question Answering) aims at question answering through reading text information images. It differentiates from the original VQA as requires large amounts of scene-text relationship addition to cross-modal grounding capability. In this paper, we propose Localize, Group, and Select (LOGOS), a novel model which attempts tackle problem multiple aspects. LOGOS leverages two tasks better localize key image, utilizes scene...
Data is of vital importance in the development machine learning technologies. Recently, within information retrieval field, a number neural ranking frameworks have been proposed to address ad-hoc search. These models usually need large amount query-document relevance judgments for training. However, obtaining this kind needs lot money and manual effort. To shed light on problem, researchers seek use implicit feedback from users search engines improve performance. In paper, we present new...
We propose a method to automatically detect 3D poses of closely interactive humans from sparse multi-view images at one time instance. It is challenging problem due the strong partial occlusion and truncation between no tracking process provide priori information. To solve this problem, we first obtain 2D joints in every image using OpenPose human semantic segmentation results Mask R-CNN. With triangulated joints, two-stage assembling proposed select correct pose thousands seeds combined by...
Open-domain Keyphrase extraction (KPE) on the Web is a fundamental yet complex NLP task with wide range of practical applications within field Information Retrieval. In contrast to other document types, web page designs are intended for easy navigation and information finding. Effective encode layout formatting signals that point where important can be found. this work, we propose modeling approach leverages these multi-modal aid in KPE task. particular, leverage both lexical visual features...
Since the distance attenuation and strong noise, Infrared Radiation (IR) dim target detection tracking is challenging in recent years. Under this circumstance, conventional particle filter track-before-detect (PF-TBD) algorithm cannot detect track effectively. In paper, a feasible two-layer based proposed for problem. The can overcome shortcomings of (PF) algorithm. By introducing local swarm reset method optimization (PSO) algorithm, it suitable low-observable multi-target tracking, has...
In order to solve the multiple waypoints path planning problem in smart home environment, we adapt optimal sampling-based algorithm (RRT*) [1] deal with navigation, namely Multi-RRT*. Our method constructs trees from waypoints, and these use simple extension connection strategy. When all are merged a single tree, an traversal will be found. Along this path, mobile robot can visit one by one. We evaluate on designed benchmark scenario compare basic bias RRT*. Finally, apply proposed...
The purpose of this study was to compare the clinical influence immediate individualized CAD/CAM healing abutments and conventional on peri-implant soft hard tissue in shaping emergence profile.Patients with a single maxillary incisor missing who accepted dental implantation were registered study. After implantation, regular prefabricated randomly inserted shape profile. A radiograph taken, pink esthetic score, papilla height, proportion, probing depth recorded at 6 months after implant...
Voice communication is the main part of wireless communication. ZigBee a new network technology, which includes low-rate, low-cost, low-energy consumption, and short distance. Although it not designed for voice communication, its 250kbps bandwidth enough to support In this paper, we achieve real-time based on AMBE-1000, CC2530, TI-Zstack protocol. We have five nodes, one coordinator, end device, other three are routers. These nodes constitute system. The system can be used in underground...
There are some problems in the traditional warehousing monitoring systems, such as complex wiring and high power consumption. Sometimes is so remote that it not quite convenient to supply alternating current with cable. This paper introduces a new environment monitor system based on wireless sensor network (WSN), which can acquire real-time parameters reduce unnecessary loss caused by emergency fire. We adopt CC2530 data transceiver, SHT11 temperature humidity sensors realize gathering of...
A number of deep neural networks have been proposed to improve the performance document ranking in information retrieval studies. However, training processes these models usually need a large scale labeled data, leading data shortage becoming major hindrance improvement models' performances. Recently, several weakly supervised methods address this challenge with help heuristics or users' interaction Search Engine Result Pages (SERPs) generate weak relevance labels. In work, we adopt two...
Most multi-view based human pose estimation techniques assume the cameras are fixed. While in dynamic scenes, should be able to move and seek best views avoid occlusions extract 3D information of target collaboratively. In this paper, we address problem online view selection for a fixed number estimate multi-person poses actively. The proposed method exploits distributed multi-agent deep reinforcement learning framework, where each camera is modeled as an agent, optimize action all cameras....