WHAT IT DOES

Click an image
Get the Air Quality

WHAT MAKES IT DIFFERENT

Dynamic Dataset of nearly 4K images

To enable further research into this field, we opensource the dataset consisting of nearly 4k HDR and Non-HDR images taken across more than 80 locations of the entire National Capital, Delhi.

On-Device Computation

Complete privacy of the user's data is ensured as all the computations are carried out On-Device and no user data can be isolated and tracked back to the user.

Generalised Model

The application uses HDR images to estimate the AQI from an image, thus nullifying the effect of the custom post processing done by each phone in the market.This ensures accurate and uniform predictions across multiple phones.

Dynamic Dataset mapped with Weather and Historic AQI Data

VisionAir

The dataset is divided into 2 sections: One contains 4K images consisting of Non-HDR images taken across multiple smartphones and 80+ locations. This dataset is useful to those looking to train a model for accurate estimates of new scenes. The other dataset contains nearly 1K HDR images taken across multiple phones extremely useful for those who wish to train a model that has high accuarcy in both scene and hardware generalisation.

Dataset Collection Application

An android Application that clicks a HDR and Non-HDR image every 15 minutes may be used by researchers to curate their own dataset.This app shall be released soon.

Privacy Preserving Mobile Application

VisionAir

To Be Released Soon!

Demo of Beta Project

For the demo of the Beta Version, Please visit Beta Project

Data Collection App

For the demo of the Data collection app, Please visit Data Collection App

Federated Learning to constantly improve Model Performance

The application is currently being tested with federated deployement to enhance the global model's performance without using any private user data.

See how to implement Federated Learning in Android here

Under Testing and Evaluation with a set of 30 active users.

Acess the CodeBase

A comprehensive codebase on github provides you access to the binary files of features, labels, and the models which may be readily used for analysis without any preprocessing.

Deep Learning

Shallow Neural Net Implementation

Find the implementation of both the Keras and Non Keras versions of the neural networks implementation at the opensourced github repository

Android

Apps to be realeased soon!