Naitri Rajyaguru

I'm a Ph.D. student in Computer Science at University of Maryland, advised by Prof. Yiannis Aloimonos in the Perception and Robotics Group.

My research focuses on developing minimal cognitive architectures by leveraging fundamental principles of AI, computational imaging, and computer vision. I aim to enable small, resource-constrained robots to perform complex tasks efficiently.

My current work addresses these core areas:

Robot Perception and Reasoning:Designing frameworks that maximize perception-driven decision-making, mirroring biological systems. Currently integrated Vision-Language Models (VLMs) as priors to enhance spatial and contextual awareness.
Minimal Data Acquisition: Utilizing computational imaging techniques to encode information directly during sensing—leveraging coded apertures, time-of-flight sensing, and polarization cameras to optimize perception with minimal data.

Previously, I worked at Zupt LLC on underwater pose estimation and in the Perception and Robotics Group on drone navigation and 3D vision.

Email / CV / Twitter / Github / LinkedIn


Ph.D. CS UMD July 2024 - Current	Computer Vision Research Engineer Zupt LLC July 2023 - Jan 2024	Research Intern Ford Summer 2022	M.Eng Robotics University of Maryland, College Park 2021-2023	Research Associate Swayyatt Robots Feb 2021 - July 2021

News

09/23/2024: Our paper "CodedVO: Coded Visual Odometry" will be presented at ICRA40 2024.
05/30/2024: Our paper "CodedVO: Coded Visual Odometry" has been accepted by RA-L 2024.

Publications

CodedVO: Coded Visual Odometry

Naitri Rajyaguru*, Sachin Shah*, Chahat Deep Singh, Cornelia Fermüller, Christopher Metzler†, Yiannis Aloimonos†

Accepted in RAL / Presented in ICRA40

CodedVO, a novel monocular visual odometry method that leverages optical constraints from coded apertures to resolve scale ambiguity of monocular depth estimation.

PDF Website arXiv

Selected Projects

Structure From Motion

3D reconstruction of a scene and camera pose estimation.

Report Code

Lottery Ticket Hypothesis in Low Data Regime

Optimized model precision through Iterative Magnitude Pruning on 5% of weights with a 1000-sample dataset, creating a generalizable model for diverse computer vision applications.

Report Code

Point Painting: Point Cloud Object Segmentation

Semantic segmentation for LIDAR & Camera using SegFormer.

Code

Super Pixel Generation using SLIC and Image Segmentation

Image segmentation using superpixels generated with SLIC and k-means with high accuracy using VGG16.

Code

Panorama Stitching

Panoramic image stitching using classical and deep learning methods.

Report Code

Website Template Credit Link.