Hyeonho OH

Research Experience

University of Southern California

At the GLAMOR Lab, our research aims to advance robots’ perception, reasoning, and control capabilities. I’m developing a simulation testing suite to evaluate robot learning policies.

At the Lira Lab, we develop algorithms for robot learning, safe and efficient human-robot interaction, and multi-agent systems. I’m developing a robust VLA model for imitation learning that can adapt to diverse environments.

Electronics and Telecommunications Research Institute (ETRI)

Led a project focused on banner detection and text recognition affected by distortion, occlusion, and noise.
Analyzed large-scale language and image data from sources including AI-Hub, ICDAR, MJ, ST, KoBERT.
Identified a decrease in model accuracy during simultaneous multi-language training (English/Korean/Chinese).
Built an multi-lanuguge dataset and developed a banner detection model based on TextBPN and CRAFT.
Constructed a contrastive learning structure to distinguish similar texts; applied CLIP and prompt tuning for inference/correction.
Contributed to IEIE publications/posters; continuing follow-up research and paper writing.

Visual AI Lab, Gachon University

Analyzed human states and fine-grained behaviors in videos for comprehensive situational understanding.
Developed a Video Swin Transformer model for behavior classification using VIRAT dataset.
Performed augmentation/transforms on Stanford40 & VIRAT with PyTorch/OpenCV; compared models with Grad-CAM, Weights & Biases (wandb), Matplotlib.
Held weekly seminars to review papers and discuss progress with researchers and the professor.

CVPR'23

Enhanced the performance of OFA and mPLUG zero-shot captioning models, achieving improved accuracy on evaluation datasets, including COCOcaptions, Flickr30k, Conceptual Captions (CC3M), and LAION
Improved the model score using prompt learning and conducted experiments to compare captioning models
Refined base model predictions by feeding generated sentences into a secondary model to enhance accuracy
Utilized prompt tuning to guide the CLIP model in identifying appropriate words

Korean Ministry of Science and ICT

Increased OCR generalization performance through an improved TRBA model
Improved the ResNet-based feature extractor with a SENet-based architecture
Applied sampling based on character frequency and employed ensemble methods to increase performance
Identified weak characters using Grad-CAM, built a focused dataset and improved accuracy through fine-tuning

Gachon University

Predicted multiple matches in the Qatar World Cup, including Argentina's win, using match records and player stats-based machine learning algorithms
Combined and sampled datasets containing player attributes, match statistics, and player evaluations for comprehensive analysis
Visualized data and results using Matplotlib, scatter plots, and correlation maps

Gachon University

Built models to detect scene events between people and objects based on YOLO, FairMOT, and HOTR
Utilized YOLOv5 and FairMOT for real-time multi-person and object detection
Created a video dataset representing interactions between people and objects
Employed HOTR (Human-Object Interaction Transformer) to detect interactions
Constructed a custom dataset and implemented server/client structure