Similar Image search using K-means clustering in Hadoop
- Given an image, the system returns k closest similar images for a database of 1 million images from Flickr.
- Used feature vectors obtained from VGG-16 network to cluster 1 million images for product quantization.
- Reduced distance computations to 0.4% of original using product quantization.
- Searches for upto top 20 images in 1 sec.
- Transformed and filtered data and implemented distributed k means clustering using Map Reduce in Hadoop.
- Designed a RESTful web service having the functionality of a typical social network; signup, login, friends and wall using Springboot framework in Java and deployed the app on Heroku.
- Used PostgreSQL database and implemented database migrations using Flyway .
- Documented the API using Swagger
- Used Maven as a build tool and Bitbucket Pipelines for CI/CD and followed Agile development practices in a team of 4.
- Created a deep reinforcement learning agent that learned to play the Flappy bird game from raw video input.
- Implemented the CNN in Python using Tensorflow, OpenCV and pygame environment.
- Studied the effect of reward structures and epsilon probabilities on the learning process.
- Based on the paper “Playing Atari with Deep Reinforcement Learning” by DeepMind.
Visualising Convolutional Neural Networks
- Visualized the effect of every filter on the input pixel space of a ConvNet using deconvolutional architecture.
- Created a custom CIFAR-100 image classifier with 67% accuracy for testing the deconvolutional net.
- Implemented un-maxpooling layers and deconvolutional layers based on Matthew Zeiler’s 2010 CVPR paper on Deconvolutional Nets and 2013 paper on Visualizing Neural Networks.
- This application visualises the camera position and orientation in the real world. As inputs you need to give in an image whose camera location we need to estimate, the original image and its dimensions.
- Uses pinhole camera model’s intrinc matrix from photogammetry
- Successfully designed and prototyped a smart IoT hub for wireless controlling of connected devices using motion gestures.
- Used Arduino Nano as the embedded system on the device with accelerometers as actuators.
- Accelerometer readings are sent to a central hub to trigger API calls to IFTTT server to controll smart lights and other devices.
- Fabricated and 3d printed the casing for the cube.
Distributed RESTful WebService
- Designed a distributed RESTful web service that emulated a busy ski-resort’s functionalities with CRUD operations
- Supported horizontal scaling by using AWS EC2 auto scaling groups. Deployed AWS RDS instances for relational data store.
- Created a rating system trained on data from Glassdoor reviews to predict company ratings based on textual reviews by using neural networks using Keras, pandas and sklearn.
- Cleaned and processed raw data to remove biases and improve result without the use of recurrent networks.
- Created a web app that simulated bitcoin trading leveraging real-time bitcoin data using Coinbase API for MHacks X Hackathon 2017.
- Used the Flask micro-framework in Python with SQLite. Used Plot.ly to create statistical visualizations. Created the application in 30hrs with a team of 4.
ArcGIS tool for locating access points for optimum WiFi signal coverage
- Developed an ArcGIS tool for identification of WiFi access points for optimum signal coverage for any given input map image of an area in vector format.
- Collected field data, conducted elevation surveys and gathered satellite imagery of the experimental area.
- Processed all the data using the ArcGIS suite.
- Worked on the programming and scripting of the tool in Python using the arcpy library
Use of remote sensing in water runoff estimation using SCS-CN method
- Used satellite imagery to estimate water runoff from study area (Anandwan) with the help of SCS-CN method in ArcGIS.
- Used Python scripts to compute quantities and generate representative graphs.