Zheng Wu

ORCID: 0000-0001-7147-173X
Publications
Citations
Views
---
Saved
---
About
Contact & Profiles
Research Areas
  • Advanced Neural Network Applications
  • Face and Expression Recognition
  • Reinforcement Learning in Robotics
  • Image Retrieval and Classification Techniques
  • Robotics and Sensor-Based Localization
  • Neural Networks and Applications
  • Robotic Path Planning Algorithms
  • Gait Recognition and Analysis
  • Autonomous Vehicle Technology and Safety
  • Human Pose and Action Recognition
  • Advanced Image and Video Retrieval Techniques
  • Hand Gesture Recognition Systems
  • Multimodal Machine Learning Applications
  • Face recognition and analysis
  • Modular Robots and Swarm Intelligence
  • Domain Adaptation and Few-Shot Learning
  • Traffic control and management
  • Infrastructure Maintenance and Monitoring
  • AI in cancer detection
  • Digital Imaging for Blood Diseases
  • Video Surveillance and Tracking Methods
  • Medical Image Segmentation Techniques
  • Anomaly Detection Techniques and Applications
  • Automated Road and Building Extraction
  • Interactive and Immersive Displays

Fuzhou University
2021-2024

Pearl River Hydraulic Research Institute
2024

University of Hong Kong
2022

University of Science and Technology of China
2019-2021

Chinese University of Hong Kong
2020-2021

University of California, Berkeley
2020

University of Chinese Academy of Sciences
2018-2019

First Affiliated Hospital of Xi'an Jiaotong University
2019

Shanghai Jiao Tong University
2006-2018

Massachusetts Institute of Technology
2018

We present Self-Ensembling Single-Stage object Detector (SE-SSD) for accurate and efficient 3D detection in outdoor point clouds. Our key focus is on exploiting both soft hard targets with our formulated constraints to jointly optimize the model, without introducing extra computation inference. Specifically, SE-SSD contains a pair of teacher student SSDs, which we design an effective IoU-based matching strategy filter from formulate consistency loss align predictions them. Also, maximize...

10.1109/cvpr46437.2021.01426 article EN 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2021-06-01

Existing single-stage detectors for locating objects in point clouds often treat object localization and category classification as separate tasks, so the accuracy confidence may not well align. To address this issue, we present a new detector named Confident IoU-Aware Single-Stage Detector (CIA-SSD). First, design lightweight Spatial-Semantic Feature Aggregation module to adaptively fuse high-level abstract semantic features low-level spatial accurate predictions of bounding boxes...

10.1609/aaai.v35i4.16470 article EN Proceedings of the AAAI Conference on Artificial Intelligence 2021-05-18

With the fast development of effective and low-cost human skeleton capture systems, skeleton-based action recognition has attracted much attention recently. Most existing methods use Convolutional Neural Network (CNN) Recurrent (RNN) to extract spatio-temporal information embedded in sequences for recognition. However, these approaches are limited ability relational modeling a single skeleton, due loss important structural when converting raw data adapt input format CNN or RNN. In this...

10.1109/icme.2019.00147 article EN 2022 IEEE International Conference on Multimedia and Expo (ICME) 2019-07-01

In the past decades, we have witnessed significant progress in domain of autonomous driving. Advanced techniques based on optimization and reinforcement learning become increasingly powerful when solving forward problem: given designed reward/cost functions, how should optimize them obtain driving policies that interact with environment safely efficiently. Such has raised another equally important question: what optimize? Instead manually specifying reward it is desired can extract human...

10.1109/lra.2020.3005126 article EN IEEE Robotics and Automation Letters 2020-06-25

This paper presents a new approach to boost single-modality (LiDAR) 3D object detector by teaching it sim-ulate features and responses that follow multi-modality (LiDAR-image) detector. The needs LiDAR-image data only when training the detector, once well-trained, LiDAR at inference. We design novel framework realize approach: re-sponse distillation focus on crucial response samples avoid most background samples; sparse-voxel learn voxel semantics relations from esti-mated voxels;...

10.1109/cvpr52688.2022.01327 article EN 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2022-06-01

The efficient planning of stacking boxes, especially in the online setting where sequence item arrivals is unpredictable, remains a critical challenge modern warehouse and logistics management. Existing solutions often address box size variations, but overlook their intrinsic physical properties, such as density rigidity, which are crucial for real-world applications. We use reinforcement learning (RL) to solve this problem by employing action space masking direct RL policy toward valid...

10.48550/arxiv.2502.13443 preprint EN arXiv (Cornell University) 2025-02-19

Existing single-stage detectors for locating objects in point clouds often treat object localization and category classification as separate tasks, so the accuracy confidence may not well align. To address this issue, we present a new detector named Confident IoU-Aware Single-Stage Detector (CIA-SSD). First, design lightweight Spatial-Semantic Feature Aggregation module to adaptively fuse high-level abstract semantic features low-level spatial accurate predictions of bounding boxes...

10.48550/arxiv.2012.03015 preprint EN other-oa arXiv (Cornell University) 2020-01-01

We always use CAPTCHA(Completely Automated Public Turing test to Tell Computers and Humans Apart) prevent automated bot for data entry. Although there are various kinds of CAPTCHAs, text-based scheme is still applied most widely, because it one the convenient user-friendly way daily user [1]. The fact that segmentations different types CAPTCHAs not same, which means CAPTCHA's bottleneck segmentation. Once we could accurately split character, problem be solved much easier. Unfortunately, best...

10.1109/icit.2017.7915494 article EN 2022 IEEE International Conference on Industrial Technology (ICIT) 2017-03-01

The self-organizing map (SOM) is an efficient tool for visualizing high-dimensional data. In this paper, the clustering and visualization capabilities of SOM, especially in analysis textual data, i.e., document collections, are reviewed further developed. A novel approach based on SOM proposed task text mining. first transforms space into a multidimensional vector by means encoding. Afterwards, growing hierarchical (GHSOM) trained used as baseline structure to automatically produce maps with...

10.1109/tnn.2007.905858 article EN IEEE Transactions on Neural Networks 2008-02-01

This article introduces a secure tunnel fast marching tree motion planning algorithm (ST-FMT*) to provide and optimal path quickly for mobile robot. The proposed ST-FMT* consists of preprocessing exploring procedures, which are responsible establishing optimizing the path, respectively. In process, generalized Voronoi graph is adopted build an equidistant roadmap generates initial collision-free solution rapidly. Then, established via minimum distance from obstacles facilitate concentration...

10.1109/tie.2021.3075852 article EN IEEE Transactions on Industrial Electronics 2021-05-03

Finding a new position for each landmark is crucial step in active shape model (ASM). Mahalanobis distance minimization used this finding, provided there are enough training data such that the grey-level profiles follow multivariate Gaussian distribution. However, condition could not be satisfied most cases. In paper, method support vector machine (SVM) based ASM (SVMBASM) proposed. It approaches finding task as small sample size classification problem, and uses SVM classifier to deal with...

10.1109/cit.2008.4594695 article EN 2008-07-01

Accurate nuclei segmentation plays an essential role in medical research and various clinical applications. Recently, deep learning has demonstrated its superior performance on object natural scene images. However, these methods cannot produce fine histopathological Therefore, this paper improves Mask R-CNN for which is called Nuclei R-CNN, mainly focus the network model, training scheme preprocess of data. Additionally, we found poor prediction accuracy under high-resolution images proposed...

10.1109/icicsp48821.2019.8958541 article EN 2019-09-01

Automatic pancreas segmentation with high precision in Computed Tomography (CT) images is a fundamental issue both medical image analysis and computer-aided diagnosis (CAD). However, challenging because of the variability location anatomy organs, while occupying only very small part entire abdominal CT scans. Due to rapid development CAD system urgent need for clinical treatment, demanded. In this paper, we propose new approach automatic using inter-/intra-slice contextual information...

10.1109/embc.2019.8856774 article EN 2019-07-01

Mobile robots generally work in harsh and restricted environments, which poses challenges for mobile to find a feasible path efficiently. This article presents planning method, namely, sampling-enhanced exploration tree (SET), improve computational efficiency environments while guaranteeing high-quality performance. The core of SET is exploration, consists critical areas identification, guiding-exploration, rectifying-exploration. In the identification phase, are identified based on...

10.1109/tii.2023.3240935 article EN IEEE Transactions on Industrial Informatics 2023-01-31

In human-robot interaction (HRI) systems, such as autonomous vehicles, understanding and representing human behavior are important. Human is naturally rich diverse. Cost/reward learning, an efficient way to learn represent behavior, has been successfully applied in many domains. Most of traditional inverse reinforcement learning (IRL) algorithms, however, cannot adequately capture the diversity since they assume that all a given dataset generated by single cost function. this paper, we...

10.1109/iros45743.2020.9341371 article EN 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2020-10-24

A method is introduced for mapping documents, based on document citations, a two dimensional map clustering and visualization the application of technology forecasting. The citation data used to build an adjacency matrix which describes set as undirected graph. dimensionality reduced using principal components analysis. dimension train small rectangular self organizing (SOM). After training, each document's input vector premultiplied by SOM weight find spatial response across centroid this...

10.1109/ijcnn.2001.938456 article EN 2002-11-13

To boost a detector for single-frame 3D object detection, we present new approach to train it simulate features and responses following trained on multi-frame point clouds. Our needs clouds only when training the detector, once trained, can detect objects with as inputs during inference. For this purpose, design novel Simulated Multi-Frame Single-Stage Detector (SMF-SSD) framework: multi-view dense fusion densify ground-truth generate cloud; self-attention voxel distillation facilitate...

10.1145/3503161.3548123 article EN Proceedings of the 30th ACM International Conference on Multimedia 2022-10-10

The Self-Organizing Map (SOM) is an efficient tool for visualizing high-dimensional data. In this paper, intuitive and effective SOM projection method proposed mapping data onto the two-dimensional grid structure with a growing self-organizing mechanism. learning phase, trained cell used as baseline framework. ordination new to map input vector so that mapped of without having plot weight values, resulting in easy visualization demonstrated on four different sets, including 118 patent set...

10.1142/s0129065703001662 article EN International Journal of Neural Systems 2003-10-01

We propose a novel approach for instance segmentation given an image of homogeneous object cluster (HOC). Our learning is one-shot because single video captured and it requires no human annotation. intuition that images objects can be effectively synthesized based on structure illumination priors derived from real images. A solver proposed iteratively maximizes our structured likelihood to generate realistic HOC. Illumination transformation scheme applied make the synthetic share same...

10.24963/ijcai.2018/144 article EN 2018-07-01

10.1016/j.engappai.2010.09.001 article EN Engineering Applications of Artificial Intelligence 2010-09-27

We propose a novel approach for instance segmen- tation given an image of homogeneous object clus- ter (HOC). Our learning is one-shot be- cause single video cap- tured and it requires no human annotation. in- tuition that images objects can be effectively synthesized based on structure illumination priors derived from real images. A solver proposed iteratively maximizes our structured likelihood to generate realistic im- ages HOC. Illumination transformation scheme applied make the...

10.48550/arxiv.1802.00383 preprint EN other-oa arXiv (Cornell University) 2018-01-01
Coming Soon ...