Projects

I work at the intersection of AI, robotics, and computer vision, developing practical solutions across funded projects, consultancies, course projects, and applied research, with a strong focus on efficient inference pipelines and scalable ML processing stacks that bridge innovative algorithms with deployable, real-world systems.

Aerial Imagery Map CAB
Project Deliverables Computer Vision Analytics for Corrosion Detection and Thermography.
  • Developing deep learning–based vision pipelines for automated corrosion detection and thermal anomaly analysis in industrial inspections.
Sports Video Analytics
Consultancy Computer Vision Analytics for Sports Video Understanding
  • Developing deep learning–based computer vision pipelines for sports analytics in fast-paced and dynamic environments.
  • Implemented racquet detection, shuttle tracking, hit detection, and player tracking.
  • Designed optimized inference pipelines for large-scale video data.
  • Improved end-to-end inference efficiency while maintaining detection and tracking accuracy.
Aerial Imagery Map CAB
Project Deliverables Sim-to-Real Generalization for Deep Learning Models for Aerial Imagery.
  • Led the AI perception component of a 3-year collaborative project (2022–2025), evaluating state-of-the-art deep learning models for aerial imagery analysis.
  • Assessed SOTA models for object detection, semantic segmentation, change detection, and continual learning under realistic Sim-to-Real conditions.
  • Evaluated advanced object detection models analyzing robustness across varying slant angles and weather conditions.
  • Generated fused RGB–IR and RGB–SAR visualizations using deep learning models.
  • Benchmarked proposed DeepMAO for segmentation task for autonomous and land classification datasets, demonstrating superior performance of overcomplete architectures.
Project Deliverables Data Generation Using Generative Modeling for Side-Scan Imagery.
  • Addressed the challenge of limited real-world side-scan sonar (SSS) data for deep learning–based underwater object detection and segmentation.
  • Proposed a synthetic side-scan sonar dataset generated via diffusion-based generative modeling.
  • Demonstrates significant improvements in downstream tasks, including object detection and semantic segmentation.
Unsupervised Domain Adaptation
Course Project Unsupervised Domain Adaptation
  • Addressed performance degradation caused by domain shifts by enabling learning on unlabeled target data through unsupervised domain adaptation.
  • Implemented a Mean Teacher–based self-ensembling framework to improve robustness and cross-domain generalization.
  • Performed cross-domain adaptation experiments between MNIST and SVHN using labeled source data and unlabeled target data.
  • Incorporated confidence thresholding and extensive data augmentation to stabilize training and enhance learning.
  • Achieved near-supervised performance, demonstrating the effectiveness of self-ensembling strategies for domain transfer.
Robotic Manipulation Result 1 Robotic Manipulation Result 2
Course Project One-Shot Object Localization for Robotic Manipulation
  • Addressed the challenge of localizing novel objects from a single visual cue without prior object-specific training.
  • Enabled robotic pick-and-place in unstructured environments through vision-based one-shot localization.
  • Implemented a Siamese CNN with shared weights to learn similarity between cue and scene images.
  • Designed a spatial attention mechanism to extract object-specific features from the cue image.
  • Generated similarity score maps and applied SoftArgMax to accurately predict object coordinates.
  • Integrated vision-based localization with inverse kinematics control for robotic picking in both simulation and real hardware.