Projects

Built SfM/MVS pipeline using pycolmap for 3D scene reconstruction from multi-view imagery with robust feature matching and sparse-to-dense reconstruction.
Implemented and compared advanced feature detection and matching techniques including SuperPoint, SuperGlue, SIFT, and ROMA for robust correspondence and accurate camera poses.
Applied Gaussian Splatting for high-quality novel view synthesis and efficient 3D scene representation.

Other team members: Adyant Srinivasan

Report

Built end-to-end transformer-based architecture for panoptic segmentation on nuScenes dataset, extending MaskPLS with cross-modal attention mechanisms.
Integrated image and point cloud features using modified transformer decoder for unified multimodal understanding.
Investigated and implemented four positional encoding strategies to spatially align and fuse multimodal features for improved segmentation accuracy.

Other Team Members: Nivisha Gandhi

Code

Developed a UAV-based vision pipeline to automate building inspection and estimate seismic structural parameters using deep learning (LEDNet for segmentation, DETIC for object detection).
Automated estimation of building distances, plan-shape, rooftop layout, and crack detection from aerial imagery for rapid structural assessment.

Other Team Members: Jayakanth Kumar

Code / Docs

Implemented and compared 3D SLAM algorithms including LOAM, LegoLoam and LIO-SAM (LiDAR-Inertial-Odometry-SLAM) with sensor fusion of LiDAR, IMU, and GPS data using ROS framework.
Explored NDT (Normal Distributions Transform) based SLAM and localization, working with RSLIDAR and Ouster LiDAR for localization in warehouse and outdoor environments.
Worked on bundle adjustment and loop closure detection techniques to improve trajectory consistency and reduce pose estimation drift over long trajectories.
Integrated Removert (Removing Moving Objects) for dynamic obstacle removal to improve SLAM in environments with moving objects and transient features.

Implemented and compared three major SLAM algorithms (Gmapping, Hector-SLAM, and Cartographer) across multiple environments using ROS ecosystem.
Evaluated performance through systematic experiments in Gazebo; analyzed mapping accuracy, loop closure detection, and pose estimation error (RMS).
Found Cartographer's better performance in complex environments and long corridors compared to traditional Gmapping and Hector-SLAM approaches.

Report

Implemented VINS-Mono and VINS-Fusion for monocular camera with IMU and camera odometry using ROS framework.
Deployed and tested both systems in warehouse environments to evaluate robustness and drift correction.

Formulated registration as a convex optimization problem leveraging Semidefinite Programming (SDP) and ADMM to achieve globally optimal and computationally efficient solutions.
Developed algorithms for both rigid and non-rigid point cloud registration with theoretical guarantees on convergence and solution quality.

Other team members: Dharmesh

Slides / Report

Integrated kinodynamic path searching and B-spline trajectory optimization into ROS2's asynchronous framework for real-time, smooth trajectory generation in autonomous drone navigation.
Ported the FastPlanner motion planning package from ROS1 to ROS2, enabling efficient trajectory generation and replanning.

Other Team Members: Basvaraj PN, Rohit Pawar

Code

Contributed to BisQue, a web-based platform providing researchers with organizational and quantitative analysis tools for up to 5D image data with flexible metadata facility.
Deployed and maintained infrastructure using Kubernetes (K3s) and Docker containerization, enabling execution of multiple deep learning based modules on distributed systems.
Developed segmentation/Object detection modules using existing deep learning algorithms(Yolo, DETR, SAM, SAM2, CellPose3D) for multimodal data analysis (3D Cell Segmentation, Underwater Image Segmentation).

Team Members: Connor, Chandrakanth, Amil Khan

Project / Code

Configured ROS Navigation Stack with obstacle avoidance, costmaps, and waypoint follower for autonomous mobile robot navigation.
Integrated Gmapping for real-time mapping and localization in indoor environments using ROS framework.
Simulated and visualized navigation system in Gazebo simulation environment with RViz for trajectory monitoring and debugging.

Built Ant Bot from scratch using Firebird V robot kit and Raspberry Pi with SFM-based line following algorithm for autonomous navigation.
Implemented segmentation-based object detection and decision-making algorithm using ArUco marker recognition for task-specific actions.

Other team members: Pranav Kumar, Ankur Bhatia, Dhawal

Code / Poster