Generate high resolution novel views of a scene which is encoded by Neural Radiance Fields. Pytorch port of the paper
Deferred Neural Rendering
Given a set of images of an object and its corresponding UV maps, we aim to perform novel view synthesis. With the help of Neural Textures and a corresponding Neural Renderer we can produce novel views.
Given an image, we wish to generate its painting version. The method uses a cyclic loss where we take the loss on ground truth and the image generated back from the generated image. Trained with a corpus of variety of paintings.
Instance Saliency Detection
Generate instance salience maps using deep learning methods. We use multi scale networks to generate salient maps which are further tuned to generate contours.
Given a set of images, generate the caption. Implementation of “Show, Attend and Tell” which uses attention based LSTM to generate each word of the caption.
Face Mask Detection
Given a set of images, we use object detection techniques to detect people wearing masks. Used Efficient Det to get a score of 0.414 mAP on mask dataset provided by AICrowd.
Neural Style Transfer
Transfering painting styles of famous painters like Van Gogh onto personal images using an optimization based technique. Gatys,2015
Computer Vision Toolbox
Implemented Chrome Keying, Image Stitching, Stereo Matching,GrabCut, Image Classification using CNNs and Optical Flow in Python
Given a set of captcha, predict the sum of digits. Used classical methods like connected components to separated out the numbers and a DL based classifier to predict the number.
Implementing a basic ray tracer in C++. Can handle volume rendering and perin noise and simple geometries.
3D airplane simulator built in OpenGL. Can view with different camera angles, perform barrels, and shoot bullets
Clone developed in WebGl. Features include noir mode and multiple levels
Multimodal Summarization of News Articles
Generate high quality summarizations of news articles by taking text and image modalities and leveraging visio-linguistic transformers like OSCAR and UNITER.
Given a textual description, we aim to generate images. The method uses a cyclic loss where we take the loss on the ground truth text and text generated on the generated image. Tested with a novel Stanford Dogs with Captions dataset.
Understanding the discussion topics and trends during the “#METoo period” in 2017 using unsupervised techniques like Topic Modeling
SLAM: Pose Graph Estimation
Given estimate of robot’s trajectory, we use non linear optimization to improve our estimate. Also referred to as SLAM problem.
Generating a dense 3D point cloud reconstruction of a scene from stereo images and synthesize a new image taken by a virtualmonocular camera using PnP algorithm
Extreme Tic Tac Toe
AI Bot which plays Extreme Tic Tac Toe, a variant of Ultimate Tic Toe. Finished 15/80 in the batch tournament with other agents. Used different heuristics combined with Min Max alpha beta pruning algorithm.
Molecular Dynamics Prediction
Learning molecular dynamics with simple language model built upon long short-term memory neural network
Portal to conduct quizzes. Admin can create different types of quizzes and users play them, Backend in Go(Gin) and Frontend in React
News streaming app which accumulates news where the user can comment and react as well. Made in Flask.
Twitter clone built with Ruby in Rails. Has authentication, forgotten password built in with other basic functionalities
Shell Clone developed in C++. Able to run background jobs, piping, and return process info
SQL engine developed in Python. Can perform basic queries like SELECT, AGGREGATE and JOIN
Basic Proxy Server in Python. Can block websites and allow access to certain users
Guess the name
Crowdsourcing information to fill gaps in hindi wikidata. The objective is to make connections between loosely related entities.
Space Invaders clone built in Pygame. Able to shoot multiple types of missiles and freeze enemies.