PROJECTS

ReDiPrune is training-free visual token pruning method for improving the efficiency of multimodal large language models. Our key idea is to prune tokens before the vision-language projector, where visual features are still rich and discriminative, instead of pruning after projection on compressed representations. To preserve both usefulness and coverage, we score tokens with a lightweight rule that jointly considers text-conditioned relevance and max-min diversity, so the selected tokens are both query-relevant and non-redundant. ReDiPrune is fully plug-and-play, requires no retraining or architectural changes, and can be seamlessly integrated into existing MLLM pipelines.

The precise identification and categorization of seizure stages through observable behaviors is essential for understanding and diagnosing neurological disorders. Traditional studies in neurology rely heavily on manual observation, which is both time-intensive and requires considerable manpower.

Utilizing rodent models for epilepsy research offers a pathway to discovering more effective drugs or neurotherapy treatments. The preliminary screening of potential anti-epileptic drugs (AED) with these models necessitates extensive video surveillance, a process that is not only slow and subjective but also prone to mistakes.

This research aims to introduce an AI-powered method for the quantitative analysis of rodent behaviors, including the classification of epilepsy stages. By employing advanced deep learning and computer vision technologies, we created an automated system for detecting and recognizing seizures in animals. A novel hybrid technique, merging model-based and data-driven approaches, was crafted to automate the classification of epilepsy stages from video footage.

This investigation validates the practicality of using video analysis for detecting and classifying seizure stages in rodent models of temporal lobe seizures, offering a consistent and quantitative way to study rodent behavior. This could potentially aid further behavioral studies in animals, such as those related to depression and anxiety. In the future, we plan to include and share a public dataset with the research community.

In this project, we present an AI–IoT system for real-time air pollution monitoring, prediction, and mapping to support sustainable urban planning. Our framework integrates low-cost IoT sensors that measure pollutants such as PM2.5, CO, NO₂, SO₂, and O₃ with machine learning models to classify air quality levels, detect pollution hotspots, generate alerts, and support adaptive waste management.

Our main technical contribution is a stacked ensemble model that combines XGBoost, CatBoost, and Gradient Boosting as base learners with logistic regression as the meta-learner. Using 5,999 sensor records collected across urban locations, our approach achieved the best performance among the tested models, reaching approximately 98.8% accuracy for multiclass AQI prediction and 99.85% for binary safe-versus-harmful classification.

Overall, we demonstrate that integrating sensing, cloud analytics, visualization, alerts, and waste-route optimization within a unified framework can enable real-time pollution monitoring, improve public awareness, and support smarter environmental decision-making. We also note that future work should address challenges related to scalability, sensor calibration, privacy, and large-scale deployment.

Read our research paper, AI-IoT Driven Real-Time Air Pollution Monitoring and Mapping for Sustainable Urban Planning.

UU-Mamba is an uncertainty-aware segmentation framework designed for reliable cardiovascular MRI analysis. It integrates the Mamba selective state space model into a U-Net–style architecture to capture both local anatomical features and long-range spatial dependencies with high computational efficiency. By incorporating uncertainty estimation into the training process, the model improves robustness and reliability, particularly in challenging anatomical regions, enabling more stable and clinically meaningful segmentation results.

The paper can be accessed at: UU-Mamba: Uncertainty-aware U-Mamba for Cardiac Image Segmentation.

IntelliCardiac is a full-stack medical imaging web platform designed for cardiac segmentation and disease classification using cine MRI.

Combining a deep learning-based segmentation model with a two-step classification pipeline. The segmentation module gains an overall accuracy of 92.6%. The classification module, trained on characteristics taken from segmented heart structures, achieves 98% accuracy in five categories.

These results outperform current state-of-the-art systems that integrate both segmentation and classification. Supporting real-time visualization, workflow integration, and AI-assisted diagnostics, IntelliCardiac shows strong promise as a scalable, accurate tool for clinical decision-making in cardiac imaging.

The paper can be accessed at: IntelliCardiac: An Intelligent Platform for Cardiac Image Segmentation and Classification.

Human action recognition is an important but challenging problem. With advances in low-cost sensors and real-time joint coordinate estimation algorithms, reliable 3D skeleton-based action recognition is now feasible.

We are interested in using recursive neural network (RNN) models to solve skeleton based human activity recognition problem.

The ability to detect manipulation in multimedia data is vital in digital forensics. Existing Image Manipulation Detection (IMD) methods are mainly based on detecting anomalous features arisen from image editing or double compression artifacts.

All existing IMD techniques encounter challenges when it comes to detecting small tampered regions from a large image. Moreover, compression-based IMD approaches face difficulties in cases of double compression of identical quality factors.

To investigate the State-of-The-Art (SoTA) IMD methods in those challenging conditions, we introduce a new Challenging Image Manipulation Detection (CIMD) benchmark dataset, which consists of two subsets, for evaluating editing-based and compression-based IMD methods, respectively.

The dataset images were manually taken and tampered with high-quality annotations. In addition, we propose a new two-branch network model based on HRNet that can better detect both the image-editing and compression artifacts in those challenging conditions.

Extensive experiments on the CIMD benchmark show that our model significantly outperforms SoTA IMD methods on CIMD.

Read our research paper, A New Benchmark and Model for Challenging Image Manipulation Detection.

UA-DETRAC is a challenging real-world multi-object detection and multi-object tracking benchmark. The dataset consists of 10 hours of videos captured with a Cannon EOS 550D camera at 24 different locations at Beijing and Tianjin in China.

The videos are recorded at 25 frames per seconds (fps), with resolution of 960 x 540 pixels. There are more than 140 thousand frames in the UA-DETRAC dataset and 8,250 vehicles that are manually annotated, leading to a total of 1.21 million labeled bounding boxes of objects.

We also perform benchmark tests of state-of-the-art methods in object detection and multi-object tracking, together with evaluation metrics detailed in our work.

Object tracking, which aims to ex-tract trajectories of single or multiple moving objects in a video sequence, is a crucial step in understanding and analyzing video sequences.

A robust and reliable tracking system is the basis for a wide range of practical applications ranging from video surveillance and autonomous driving to sports video analysis.

We have developed several object tracking algorithms based on the hyper-graph formalism and provided a benchmark dataset for multi-object tracking performance evaluation.