Object Detection in 360° Videos

AWL Inc., Japan
  • Worked on object detection in 360° fisheye videos with very little unlabelled data
  • Used Faster RCNN with Feature Pyramid Network (FPN) as feature extractor and explored augmentation techniques relvant to the problem on COCO dataset
  • Investigated various alternative approaches including head detection, pose detection, multi label classification for the task (details in report)

Thermal Face Recognition

Guide: Wei-Ta Chu | National Chung Cheng University, Taiwan
  • Worked on cross-modal face matching from thermal to visible spectrum
  • Formulated a method for transforming thermal to visible face images using a Residual U-Net and Pixel Shuffle for upsampling
  • Achieved state of the art performance on NU dataset in terms of recognition accuracy and visual quality of generated images

Instrumental Music Translation

Guide: Preethi Jyothi | IIT Bombay
  • Implemented a system for translating music from one instrument to another without using any paired data
  • Used a LSTM based encoder decoder with attention in which the encoder is shared across all instruments while decoders are instrument specific
  • Encoder was adversarially trained using domain confusion loss to get instrument independent intermediate representations

Automated Dispensary

Institute Technical Summer Project | IIT Bombay
  • Designed an automated dispensary mechanism to locate and identify medicines kept on a table and pass the required medicine to the user
  • Made a multi-class CNN classifier on TensorFlow using transfer learning for recognizing the pills from live images which were processed using OpenCV-Python library

FAQ Bot for Freshmen

WnCC Seasons of Code | IIT Bombay
  • Made a chatbot using Rasa NLU for answering queries of newly joined students at IIT Bombay
  • Designed a model to perform intent classification into 18 different classes, entity recognition and a mapping of intents to a set of predetermined replies

Cell Segmentation using U-Net

Guide: Suyash Awate | IIT Bombay
  • Implementated the U-Net for Image Segmentation paper in PyTorch
  • Evaluated it on Cell tracking challenge and EM Stacks datasets
  • Experimented by adding batch normalization and Dice coefficent loss achieving better results

SDN Router: Packet Classification

Guide: Ashwin Gumaste | IIT Bombay
  • Implemented the packet classification algorithm of a 400 Gb/s Software Defined Network (SDN) router in VHDL
  • Packets from multiple input ports are classified and modified according to their header data and forwarded to one of the output ports

Dynamic State Encryption Machine

Guide: Amitabha Sanyal | IIT Bombay
  • Designed a GUI-based Enigma machine using functional programming
  • Substitution Encryption is used in which the mapping changes on each input alphabet making it much more powerful than a simple Caesar cipher
  • Designed an algorithm to guess the message given the encrypted message and a known prefix in Racket

Battleship

Guide: Soumen Chakrabarti | IIT Bombay
  • Built an online version of the game of Battleship using Django
  • Implemented persistent connections using WebSockets in Django-Channels and JavaScript for real-time content, multiple parallel game sessions and in-game chats
  • Used ReactJS for dynamic update and rendering of the board on changes in game state and jQuery for handling diverse operations like selecting DOM elements for changing their state

FGSM for Generating Adversarial Examples

Self Project
  • Implemented the Fast Gradient Sign Method to generate adversarial examples for a VGG-16 network pre-trained on ImageNet dataset
  • Implemented both untargeted and targeted attacks and tested the perturbed images on various models for checking whether they generalize

Quiz Buzzer Circuit

Guide: M. B. Patil | IIT Bombay
  • Designed a Fastest Finger First Indicator circuit that detects which of the four contestants responded by clicking their button first, disabling the other buttons
  • Used CD4033 (Decade Counter) driven by IC 555 timer to display the time taken for the first response on a pair of 7-segment displays with an accuracy of 0.1 seconds

Other Projects


  • POP3 Email: Implemented a simplified POP3 client-server system with support for multiple simultanoeus clients
  • Neural Machine Translation: Implemented a French to English translator using ANKI flashcards data in Keras
  • Music Genre Classification: Extracted spectrogram features from music clips using librosa for classification
  • MNIST Digit Classifier: Designed a neural network to classify digits from MNIST dataset in Keras and PyTorch
  • Part Downloader: Wrote a Python and Bash script to download large files in parts over unreliable internet
  • CIFAR-10 Classifier: Designed and implemented a CNN to classify images in CIFAR-10 dataset on PyTorch
  • Bluetooth App Controlled Car: Made an Android application to control a car over bluetooth
  • IOT Controlled Devices: Implemented a WiFi switch using a NodeMCU ESP8266 to operate devices over internet
  • Obstacle Avoiding Bot: Built an obstacle avoiding bot using Arduino and an servo mounted Ultrasonic sensor