Signed Graph Analysis

Signed graphs provide more information than unsigned graph networks. In this project we analyze and characterize signed graphs, propose a novel way to characterize them and discover communities.

We expand the balance theory to signed social network graph analysis and propose a frustration cloud view of the signed graph. We then quantify vertex and edge int erms of frustration cloud statistics, and validate this novel social network graphs analysis approach.

Community Discovery approaches suffer for lack of scalability or reproducability, and high modularity seems to be hard to overcome. We conduct a detailed analysis in terms of efficacy, efficiency, scalability, and reproducability of existing methods, and propose frustration cloud based approach for cluster boosting for high modular signed graphs.

Unsigned Graph Analysis

We mine the ways users interact in the large social media networks to enrich content classification and community discovery.

Scalable end-to-end Twitter network data management pipeline that gathers, stores, and models rich relationships from Twitter networks, and analyzes Twitter data using a combination of graph-clustering and topic modeling techniques at scale using multiple data science methods for graph construction and tweet data processing.

The large number of users, messages, and tags makes it difficult to separate interesting and meaningful conversation threads from the spread of fake news, malicious accounts, background noise, or irrelevant trolling on Twitter. We answer if the content of Tweets be classified based on interactions alone, and when and how well can community classification predict the content class.

Analysis of Objects in Overhead Imagery and Video Streams

This research track is primarily focused on object detection, domain adaptation, unknown class discovery, multi-camera tracking, and activity detection from aerial and satellite images/ video streams.

Robust, adaptable, intuitive implementation for automated early activity warning in maritime scenarios (piracy) or smart city setup. We expand computing with words paradigm to augment training data and identify threating activities where lack of training data prohibits the use of deep learning.

Accepted Papers:

• Overhead Projection Approach For Multi-Camera Vessel Activity Recognition.
• Computing with words in maritime piracy and attack detection systems.

We expand the recent advances in deep learning to design and train algorithms to localize, identify, track, and re-identify small maritime objects under varying conditions.

Accepted Papers:

• Identifying Maritime Vessels at Multiple Levels of Descriptions using Deep Features.
• Transfer learning of deep neural networks for visual collaborative maritime asset identification.

We explore for strong backbone and proposal generator for small object detection from satellite images. We design task-specific detection pipelines for small and complex objects. Our research extends toward unsupervised domain adaptation with a large domain gap.

Accepted Papers:

• Progressive Domain Adaptation with Contrastive Learning for Object Detection in the Satellite Imagery.
• Improving the Energy Efficiency of Real-time DNN Object Detection via Compression, Transfer Learning, and Scale Prediction.
• Small Object Difficulty (SOD) Modeling for Objects Detection in Satellite Images.

Image Analysis

Iterative computer vision algorithm (based on histogram inference) that in each iteration refines segmentation masks from existing annotation boxes in maritime domain.

A web wrapper for the Boxes2Masks application.