Career Profile
I am currently working for Intel Corporation as a Technical Consultant. I’m working in the areas of Human-object interaction and scene understanding. My research interest are Computer vision and Machine learning.
Education
- Teaching Assistant
- Doing research on Bayesian deep learning and its application to computer vision and NLP.
- Interned at Persistent systems, Nagpur.
- Completed B.Engg with 7.98 CPI
Experiences
- Currently working on Human-object interaction, scene understanding and data augmentation.
- My task is to read and implement recent papers.
- I’m Responsible to come up with deep learning methodologies that will give better accuracy.
- Using Blender 2.80 to create animated videos and 3D modelling.
- Responsible for end to end delivery of android apps.
- Worked on Android sdk, Android NDK, JNI, JSON, Webview.
- Design Patterns Used : MVC, Singleton, Adapter .
- Good experience with Ad sdk’s and Materail Design.
Projects
Deformable medical image reconstruction using Bayesian deep learning (Master's thesis) - Given a distorted image the task is to reconstruct it to original image. We developed end-to-end Bayesian convolutional neural network (CNN's) consisting of downsampling and upsampling convolution layers, pooling layers and batch normalization to register the given pair of images. Lungs CT and cardiac MRI dataset were used in this experiment and got an accuracy of 96% when compared with its actual ground truth. Tools used are pytorch 1.0, python, sklearn, matplotlib.
Data augmentation using deep convolutional Generative Adversarial Network (GAN's). - Given a dataset the task is to generate more data (augment) that is similar to the given data. Used deep convolutional GAN’s to augment data. Training was done in min-max way with two CNN networks i.e, generator and discriminator where input noise vector to generator is taken from gaussian distribution. Dataset used in this experiment is Celeb-A dataset and tools used are pytorch, spyder and matplotlib.
Text translation using RNN and attention mechanism. - Given an english language text the task is to convert it into french language. Language translation from one language to another using RNN, GRU and autoencoder along with attention Weights. Used teacher forcing as a means to train the network. Dataset used in this experiment is eng-fra translation pair data. Pytorch 1.0, skimage and matplotlib libraries were used.
Image Alignment using Convolutional Autoencoder. - Given a pair of original image and distored image the task was to align the distorted image to that of original image. We developed an autoencoder network with spatial transformer module, stochastic gradient descent as optimizer. Used MNIST data set and got accuracy upto 98% between aligned and original image after transformation. Pytorch 1.0, skimage and matplotlib libraries were used.
Offline signature verification using Siamese convolutional Network - Developed end-to-end siamese CNN which is multi-input CNN architecture containing two identical subnetworks. Our architecture can take variable size input image with the help of spatial pyramid pooling layer. CEDAR dataset were used for the experiment and observed an accuracy of 78%, Pytorch 1.0, skimage and matplotlib libraries were used.
Prediction using Random Forest. - We performed data preprocessing, data wrangling, visualization, exploratory data analysis on the given dataset. we also performed Outlier detection using boxplot,cross validation, feature selection using pearson correlation coefficient, and feature importance.Random forest approach was giving better performance compared to other models with accuracy of 89.90 on test dat Libraries used sklearn, seaborn, numpy, pandas.