Shree Nath

Shree Nath Dutt

Software Engineer · Amazon

LinkedIn · Github · Google Scholar
Contact Me

About Me

Currently working as a Software Engineer II at Amazon, in the Prime Air Engineering Team.

I finished my Master's degree from Computer Science Department at Stony Brook University in 2019. Successfully defended my thesis on Salient Object Ranking.

During my time here, I worked with Prof. Dimitris Samaras and Zijun Wei in the Computer Vision Lab.

I graduated in Electronics Engineering from IIT(BHU), Varanasi in 2015. Prior to joining Stony Brook University, I worked at Flipkart Internet Pvt. Ltd. as a Software Develoment Engineer from 2015-17.

My research interests include Computer Vision, Machine Learning and Human Computer Interaction.

Work Experience

Software Engineer - Amazon (2019 - now)

Currently Working as a Software Engineer II in the Prime Air Engineering Team.
** Project details are confidential. **

Computer Vision Intern - Trademark Vision (Summer - 2018)

Developed a deep learning algorithm to generate an Image feature representation for cross-domain Image matching and ranking of design trademarks.

Worked on development and implementation of a several computer vision algorithms to convert images of trademarked objects to corresponding industrial design sketches.

Software Development Engineer - Flipkart (2015 - 2017)

Worked in Merchandising and Monetization team, developing server side applications for serving organic and monetized content on Flipkart.com mobile app and website.

Latest Projects

Salient Objects Ranking

Computer Vision Lab · Stony Brook University

Presented my Master's thesis on Ordering Salient Objects in Images. Working on the Faster RCNN architecture, we modified the Region Proposal Network to include an Object Saliency Metric. Worked with Zijun Wei, PhD candidate at Department of Computer Science in Stony Brook University.

Building on top of the Faster RCNN architecture, we explored the task of predicting Salient regions in an image, and using this information to propose effective ROIs for detection and classification of objects. We modified the Region Proposal Network to include an Object Saliency metric, which allowed our model to learn a Saliency (Importance) Score of the regions, thereby providing an order to the eventually detected objects. Our results performed better than the existing MLNet model for ranking of the Salient objects. Since our architecture additionally penalized the model for producing false positives, our model outperformed the Faster RCNN model on mean Average Precision

Publications

A Scalable and Robust Framework for Intelligent Real-time Video Surveillance

Shreenath Dutt & Ankita Kalra · ICACCI 2016

In this paper, we present an intelligent, reliable and storage-efficient video surveillance system using Apache Storm and OpenCV. As a Storm topology, we have added multiple information extraction modules that only write important content to the disk. Our topology is extensible, capable of adding novel algorithms as per the use case without affecting the existing ones.

More on Google Scholar

Past Projects

Beyond PassGAN - Using a VAEGAN to model password distribution

Fall 2018 · Stony Brook University

Improved on the current state of the art password modeling architecture, PassGAN by using VariationalAuto Encoders, with a GAN Loss. The proposed algorithm was able to match a higher number of pass-words in same number of generated words.

Transfer learning for Digits Classification on MNIST Dataset

Fall 2017 · Stony Brook University

Trained an MNIST CNN 6-layered classifier on just the digits 1, 4, 5 and 9, and then used the trained model weights of the lower 4 layers to train a classifier for the rest of the MNIST. This approach of transfer learning helped us get an accuracy of 80% by just training on upper two layers for the rest of the digits. Link

Interactive binary segmentation based on superpixels and graph-cuts

Fall 2017 · Stony Brook University

Performed semi-automatic binary segmentation based on SLIC super-pixels and graph cut. Also developed an interactive application that lets user draw foreground and background markings, to generate a max-flow graph cut based segmented image. Link

Identification of Adult Content in Images

Fall 2014 · IIT(BHU), Varanasi

Trained a Random Forest Ensemble Classifier to give a binary classification on the basis of 17 features calculated from the Region of Interests extracted from various parts of the Image. These ROIs were isolated after training a Logistic Regression Model to classify parts of Image as Skin and non skin.

Gaussian Mixture Models based Visual Servoing

Summer 2013 · IIIT Hyderabad

Implemented GMM based Visual Servoing using Dense Depth Maps. Worked on previously presented GMM based Visual Servoing for controlling 3 DOF of Motion of the Camera to generalize it for 6 Degrees of Freedom.

More on Github