CW-BASS: Confidence-Weighted Boundary Aware Learning for Semi-Supervised Semantic Segmentation
Ebenezer Tarubinga,
Jenifer Kalafatovich,
Seong-Whan Lee
IJCNN 2025
Tackled boundary blur and confirmation bias using confidence-weighted pseudo-labels, dynamic thresholding, and a boundary-aware module. Achieves 65.9% mIoU on Cityscapes with only 100 labeled images.
FARCLUSS: Fuzzy Adaptive Rebalancing and Contrastive Uncertainty Learning for Semi-Supervised Semantic Segmentation
Ebenezer Tarubinga,
Jenifer Kalafatovich,
Seong-Whan Lee
Neural Networks – Under Review
Introduced fuzzy pseudo-labeling, entropy-based weighting, adaptive class rebalancing, and lightweight contrastive regularization to improve generalization in semi-supervised semantic segmentation.
Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learning
Liang et al., CVPR 2024
Studied stealthy backdoor triggers in CLIP-style models by aligning visual and textual embeddings in the joint representation space.
Semantic-Aware Multi-Label Adversarial Attacks
Mahmood et al., CVPR 2024
Crafted targeted perturbations for multi-label classifiers that exploit semantic label co-occurrence dependencies.
Semi-Supervised Segmentation Baselines: UniMatch & ST++
Yang et al., CVPR 2022 / 2023
Systematic evaluation of weak-to-strong consistency and self-training techniques that served as the foundation for CW-BASS and FARCLUSS.
Monocular Depth Estimation for Autonomous Driving
Yang et al., NeurIPS 2024
Depth estimation and point cloud generation for autonomous driving applications using Depth Anything V2, with an improved depth-to-pointcloud pipeline.
Scalable Urban Dynamic Scenes
Turki et al., CVPR 2023
Multi-branch hash tables and NeRF-style encoding for large-scale dynamic urban scene reconstruction.
Speech Emotion Recognition
Wu et al., ICASSP 2023
Bi-GRU with self-attention for classifying emotions from audio and text signals.
Text-to-Image Generation
Various
Comparative study of autoregressive, controllable, and diffusion-based approaches to text-conditioned image generation.
Aligning Segment Anything Model to Open Context via RL
Huang et al., CVPR 2024
Automatic SAM prompting with a reinforcement learning agent for diverse segmentation tasks.
Tool-Augmented Reward Modeling (Themis)
Li et al., ICLR 2024
Integrating external APIs into reward models for more accurate and transparent decision processes.
ASD Classification with Multi-Site fMRI Data
Kunda et al., IEEE TMI 2022
Second-order functional connectivity embedding with domain adaptation for autism detection across diverse clinical sites.
Mutual Correction Framework for Semi-Supervised Medical Image Segmentation
Wang et al., CVPR 2023
Mutual correction framework for refining medical segmentation masks and mitigating confirmation bias with limited annotations.
Why Does the Effective Context Length of LLMs Fall Short?
An et al., ICLR 2025
Investigated the gap between declared and effective context length in LLMs, analyzing positional-encoding biases and the STRING shifting method.
AgriLet — AI Crop Disease Detection & Education Platform
GliTech · Full-Stack Application
AI-powered crop disease identification with severity scoring, 6 structured learning modules, community data contributions, and downloadable agricultural datasets. Built with React, TypeScript, Express, and Prisma.
eLearn — E-Learning Platform for O/A Level Exam Prep
GliTech · Full-Stack Application
20+ subjects, 70+ lessons, progress tracking with study streaks, learning analytics dashboard, and personalized settings. Built with React, TypeScript, Express, and Prisma.
3D Digital Twin Traffic Management System
Gractor · Production System
Edge AI-based real-time traffic monitoring with Three.js 3D visualization, predictive analytics, and smart city integration. Built with React, TypeScript, Three.js, and FastAPI.
