Kushagra Agrawal

Camera Systems Engineer

• Working in the camera simulator team, implementing camera ISP features such as Multi-Frame HDR, Multi-Frame Noise Reduction etc.

Full Stack Developer

• Designed and implemented CGI Vesa: a mobile based application serving as electronic access channel for parents. Enables users to apply for early childhood education and to report on the day to day care needs.
• Implemented different pages by using various ionic components defining routes for each these pages with its own lazy loading modules. Implemented features related to notifications, enrollment etc.
• Tech Stack - Java Spring Boot, Angular 7, Ionic 4

Student Trainee

• Worked with the Vision Research Team on developing a lightweight architecture for Semantic Segmentation.
• Focus was on developing a Deep Neural Network architecture that can be deployed on embedded devices such as smartphones.
• Developed a MobileNetV2 based architecture that achieved an mIoU of 81.2% on the PASCAL VOC 2012 dataset. In comparison, the state of the art achieves 84.4% mIoU using 41M parameters.

Birla Institute of Technology and Science

CGPA: 8.6/10

Army Public School, Mathura

Specialisation: Science

Army Public School, Bengdubi

Specialization: None

Deep Learning for Video Summarization

• Project aimed at comparing Deep Learning architectures to traditional video summarization techniques.
• Most recently used KMeans algorithm to cluster video frames based on features extracted via a Convolutional Neural Network pre-trained on ImageNet dataset.

Augmented Reality for Gait Analysis

• Project aimed at developing an Augmented Reality system to perform gait correction exercises in a cost effective manner.
• The application developed gamifies the exercises for correcting one’s gait.The current prototype developed studies difference in movement of subjects in a virtual obstacle course vs a real one.

Deep Learning for Image Classification

• Studied various methods for image classification.
• Developed system used Convolutional Neural Networks. It was a VGG16 architecture pre-trained on ImageNet dataset. By adjusting the final fully connected layer, an accuracy of 75% on the CIFAR-10 dataset was achieved.

Course Projects

• Computational Geometry - Implemented an optimal algorithm to find kernel in a polygon.Source Code.
• Artificial Intelligence - Compared various Machine Learning algorithms on the task of classifying text into pre - defined categories. Source code.
• Information Retrieval - Built a vector space model based Information Retrieval System. Source code.
• Machine Learning - Implemented ID3 Decision Trees, Artificial Neural Networks and Naive Bayes' algorithm.Source Code for ANN.Naive Bayes code.