Efficient Vision Transformers
Attention that scales down without giving up accuracy — partial attention, shuffled local self-attention, spatial decay matrices, and refined attention maps for recognition and dense prediction.
Open to collaboration
Xuan-Thuy Vo
Postdoctoral Researcher UNIST Vision and Learning (UVL) Lab Ulsan, South Korea
My research connects seeing, describing, predicting, and acting. I care about systems that generalize past their training data — and that run at a cost someone can actually afford.
74
Publications
13
Journal Articles
61
Conference Papers
5
Awards
2020
Publishing Since
Recent news
- Aug 2026 Joined the UNIST Vision and Learning (UVL) Lab, Graduate School of Artificial Intelligence, UNIST, as a Postdoctoral Researcher.
- 2026 Paper "Dilated Multi-Layer Perceptrons Mixer for Faster Neural Networks" published in Neural Networks.
- Jul 2025 Best Paper Award at HSI 2025 for "Efficient Multi-Scale Spatial Interactions for Visual Recognition Tasks".
- 2025 Paper "Local Self-Attention with Mixing Abstract Tokens for Urban Autonomous Driving" published in IEEE Transactions on Industrial Informatics.
- Sep 2024 "Efficient Vision Transformers with Partial Attention" presented at ECCV 2024.
- Aug 2024 Best Paper Presentation at IWIS 2024 for "Refining Attention Maps for Vision Transformer".
- Mar 2024 Started as a Postdoctoral Researcher at the Intelligent Systems Lab., University of Ulsan.
- Feb 2024 Defended my Ph.D. thesis, "Efficient Vision Transformers for Object Recognition", at the University of Ulsan.
Research interests
Open-Vocabulary Semantic Segmentation
Segmenting categories never seen at training time by grounding pixel-level representations in language rather than a closed label set.
Vision-Language Models
Joint image-text representations, cross-modal fusion, and text-guided visual learning for retrieval, re-identification, and grounded recognition.
Multimodal Large Language Models
Adapting and evaluating MLLMs — GPT-4o, LLaVA, Qwen, LLaMA, MiniCPM — for visual reasoning tasks under practical latency and memory constraints.
Autonomous AI Agents
LLM agents that plan and call tools — ReAct and code-agent loops, LangChain and LlamaIndex pipelines, and deployed voice agents.
Latent World Models
Prediction in latent rather than pixel space — object-centric models that discover keypoints, boxes, and masks from video alone, learn action spaces without annotation, and trade extra loss terms for inductive bias.
Publications
74 peer-reviewed Type
Year
Showing 74 of 74
2026
Dilated Multi-Layer Perceptrons Mixer for Faster Neural Networks
2025
ESIFCD: A Dual-Temporal Cross-Domain Fusion Network for Remote Sensing Change Detection
Optimal Proxy Mining Contrastive Network for Unsupervised Person Re-Identification
Text-guided visual representation learning via cross-modal fusion for person re-identification
FARVIS: A Facial Attribute Recognizer based on Vision Sensor using Super-Efficient Network Supporting Service Robots
A High-Accuracy and Faster Face Recognizer Supporting Biometric Continuous Authentication for Smart Factory Workers
Local Self-Attention with Mixing Abstract Tokens for Urban Autonomous Driving
Improved YOLOv11 Based on Attention and Extra Head Modules for Dental Disease Detection
Improved YOLO11-Based Weapon Detection Using Hybrid Ghost Bottleneck and Attention Module for Security Surveillance System
Self-Attention with Spatial Decay Matrix Based on Inverse Multiquadric
A Hybrid Vision Transformer and Convolutional Neural Network Architecture for Banana Leaf Disease Classification
Efficiency-Accuracy Trade-Off of Facial Attribute Classifier Supporting Human-Robot Interaction
Efficient Human Behavior Detector for Vision-based Emergency Evacuation Systems
Efficient Multi-Scale Spatial Interactions for Visual Recognition Tasks
Artificial Behavior Intelligence: Technology, Challenges, and Future Directions
Real-time facial attribute classifier using streamlined feature re-calibration module
Vision-Based Human Fall Detection with Pre-trained YOLO Models: Comparison and Application
Personal Protective Equipment Detection Based on Improved YOLO11-Nano Using Multi-Perspective Attention
2024
HFD-YOLO: Improved YOLO Network Using Efficient Attention Modules for Real-Time One-Stage Human Fall Detection
An Efficient Detector for Automatic Tomato Classification Systems
Efficient Vision Transformers with Partial Attention
Improved Tomato Detector Supporting for Automatic Harvesting Systems
Small Object Detection Without Attention for Aerial Surveillance
Multi-Scale Convolutions Meet Group Attention for Dense Prediction Tasks
Wider neighborhood-aware attention in improving YOLOv8n for one-stage human fall detection
Refining Attention Maps for Vision Transformer
Facial Attribute Recognition Using Weighted Bottleneck Transformer for Intelligent Marketing
Simple Human Fall Surveillance System Based on Person Detection
EMPCNet: Facial Attribute Recognition Using Efficient Multi-Perspective Convolution for Human-Robot Interaction
Simultaneous Facial Age Group and Gender Recognition using Efficient Local-Global Attention Network for Intelligent Advertising
Reweighting foveal visual representations
Fruit Ripeness Detector for Automatic Fruit Classification Systems
Human Facial Age Group Recognizer Using Assisted Bottleneck Transformer Encoder
Minor Object Recognition from Drone Image Sequence
2023
Car Detection for Smart Parking Systems Based on Improved YOLOv5
Categorical weighting domination for imbalanced classification with skin cancer in intelligent healthcare systems
Facial Attribute Recognition Using Lightweight Multi-Label CNN-Transformer Architecture for Intelligent Advertising
Vehicle Detector Based on Improved YOLOv5 Architecture for Traffic Management and Control Systems
A Simple Yet Effective Data Augmentation for Human Pose Estimation
Multi-Band Feature Fusion in Satellite Images for Land Cover Classification
Simultaneous Person, Face, and Hand Detector Based on Improved YOLOv5
Hierarchical Vision Transformers with Shuffled Local Self-Attentions
Gender Recognizer based on Human Face using CNN and Bottleneck Transformer Encoder
Car Detector Based on YOLOv5 for Parking Management
Unifying Local and Global Fourier Features for Image Classification
Age Group Recognizer based on Human Face Supporting Smart Digital Advertising Platforms
Human face detector with gender identification by split-based inception block and regulated attention module
YOLO5PKLot: A Parking Lot Detection Network Based on Improved YOLOv5 for Smart Parking Management System
Dynamic Circular Convolution for Image Classification
2022
Accurate Bounding Box Prediction for Single-Shot Object Detection
A Review on Anchor Assignment and Sampling Heuristics in Deep Learning-based Object Detection
Combination of deep learner network and transformer for 3D human pose estimation
A facial gender detector on CPU using multi-dilated convolution with attention modules
Balancing Multiple Object Tracking Objectives based on Learned Weighting Factors
An efficient face-based age group detector on a CPU using two perspective convolution with attention modules
Fire warning based on convolutional neural network and inception mechanism
A Study on Efficient Multi-task Networks for Multiple Object Tracking
Robust hand detection based on convolutional neural network and attention module
Efficient High-Resolution Network for Human Pose Estimation
Fusion of Segmentation and Classification for Improving Skin Disease Diagnosis
Multi-level feature reweighting and fusion for instance segmentation
High-Resolution Network with Attention Module for Human Pose Estimation
Pedestrian head detection and tracking via global vision transformer
Convolutional neural network design for eye detection under low-illumination
2021
Light-weight convolutional neural network for distracted driver classification
Regression-aware classification feature for pedestrian detection and tracking in video surveillance systems
Dynamic multi-loss weighting for multiple people tracking in video surveillance systems
Triple detector based on feature pyramid network for license plate detection and recognition system in unusual conditions
Stair-step feature pyramid networks for object detection
Efficient Spatial-Attention module for human pose estimation
2020
Simple fine-tuning attention modules for human pose estimation
3D SaccadeNet: A single-shot 3D object detector for LiDAR point clouds
Enhanced feature pyramid networks by feature aggregation module and refinement module
Bidirectional non-local networks for object detection
No publications match those filters.
Experience & education
Appointments
- Aug 2026 — PresentPostdoctoral ResearcherUNIST Vision and Learning (UVL) Lab, Graduate School of Artificial Intelligence, UNISTUlsan, South Korea
- Mar 2024 — Jun 2026Postdoctoral ResearcherIntelligent Systems Lab., University of UlsanUlsan, South Korea
- Supervisor: Kang-Hyun Jo
- Sep 2018 — Feb 2024Research AssistantUniversity of UlsanUlsan, South Korea
- Supervisor: Kang-Hyun Jo
- Jun 2018 — Aug 2018Engineering InternSamsung Electro-Mechanics VietnamVietnam
- Jun 2017 — Aug 2017Undergraduate InternTruong Giang Electric JSCDa Nang, Vietnam
- Supervisor: Nguyen Kim Anh
Education
- Sep 2018 — Feb 2024Ph.D. · Electrical, Electronic and Computer EngineeringUniversity of UlsanUlsan, South Korea
- Thesis: Efficient Vision Transformers for Object Recognition
- Supervisor: Kang-Hyun Jo
- Committee: Hee-Jun Kang, Kang-Hyun Jo, Young-Soo Suh, Hyun-Doek Kang, Moon-Ho Jeong
- Sep 2013 — Feb 2018B.Sc. · Electrical and Electronic EngineeringDa Nang University of Science and TechnologyDa Nang, Vietnam
- Cumulative GPA: 3.36 / 4.0
Awards
5 totalJul 2025
Best Paper Award
HSI 2025
Efficient Multi-Scale Spatial Interactions for Visual Recognition Tasks pdf ↗
Aug 2024
Best Presentation Paper
IWIS 2024
Refining Attention Maps for Vision Transformer pdf ↗
Aug 2023
Best Paper Award
IWIS 2023
Hierarchical Vision Transformers with Shuffled Local Self-Attentions pdf ↗
Oct 2022
Best Paper Award
MAPR 2022
Balancing Multiple Object Tracking Objectives based on Learned Weighting Factors pdf ↗
Feb 2022
BK21 Outstanding Paper Award
Brain Korea 21
Awarded for IEEE Transactions on Industrial Informatics (2022) and Neurocomputing (2022)
Professional service
Session Chair
- IWIS 2025
- HSI 2025
- ISIE 2024
- IWIS 2024
Webmaster
- ISIE 2024
- IWIS 2024
- IWIS 2023
- IWIS 2022
- IWIS 2020
Journal Reviewer
- IEEE Transactions on Industrial Informatics
- IEEE Access
- Expert Systems with Applications
Conference Reviewer
- IEEE IES Conferences — ISIE, IECON, HSI, IWIS
Technical skills
Programming
PythonPyTorchTensorFlowKerasC++MATLABJavaScript
Multimodal LLMs
GPT-4oLLaVAQwenLLaMAMiniCPMHugging Face
LLM Agents & Tool Calling
ReActReActCodeAgentLangChainLlamaIndexSmolAgent
Deployment
Voice AI Agents
Contact
I am always glad to hear about collaborations, student supervision, or reviewing invitations. Email is the fastest way to reach me.
Email xuanthuy@unist.ac.kr
Office Graduate School of Artificial Intelligence, UNIST
Address 50 UNIST-gil, Eonyang-eup, Ulju-gun, Ulsan 44919, South Korea
Profiles Google Scholar · ORCID