Verlan

Can you really count on social media posts about your lost pet? Your pet’s life shouldn’t depend on whether or not a stranger glanced at a picture of your dog, and we believe there is an alternative.

Comment

Inspiration

Dronit

Can you really count on social media posts about your lost pet? What information can you recall from the last missing poster you saw?

Our pets are way too precious for us to depend on crowd-sourced searches. Your pet’s life shouldn’t depend on whether or not a stranger glanced at a picture of your dog, and we believe there is a 21st century alternative.

Brinder

Brinder may look like a trivial fun app where you just post an image of you dog to find its breeds but it is actually a way to expand the Stanford dog dateset which is used for researching behavior psychology and diseases in dogs.

What it does

Dronit

Recently, the market for drones has surged and they are now more available than ever for professionals, students, and personal users alike. With this sudden advance in technology comes many new opportunities and applications. We, for instance, found that drones are very effective in recording daily street activity of both animals and humans in cities. We utilized this to transform drones with the help of AI to track stray and displaced dogs in cities in hopes of connecting them with their owner as well as providing a new home to some of them.

This is made possible by heavy training of a neural network with google images of pre-existing drone footage of animals, and processing everything on a google cloud server to provide maximum run-time optimization. All of this is seamlessly integrated within our user-friendly interface and provides you with the best pet-finding experience possible.

We also use reverse geocoding technology powered by google to ensure that the location of your lost dog is precise and accurate and not just a simple set of coordinates.

Easy interface: There is no looking at yucky terminal outputs or a cluttered UI. We implemented a clean, friendly design to suit everyone’s needs.

Let the tech do the searching: Our drones have been equipped with an advanced neural network powered by google cloud vision api to analyze footage and recognize your dog. You’ll receive word when your dog has been found, and you can rest peacefully at the comfort of your home knowing that drones are hard at work.

Availability: Dogs can roam around almost anywhere, but so can cheap drones with attached cameras!

Open source: Don’t trust us? Too good to be true? Want to command your own drone army to lead a search? Check out our GitHub page yourself!

Powered by 21st century tech: Our service is powered by google cloud servers, python, and flask.

Brinder

The heart and soul of this project is being able to effectively recognize dog breeds, and our team has worked meticulously to ensure that.

In fact, we were so invested in getting this right that we ended up with a world class model with a test accuracy of over 99% and a validation accuracy of over 90% with 120 different breeds. This ultra-high accurate model will ensure that you are searching for the right dog. Even larger datasets that train our model to recognize more breeds more accurately are on their way to provide you with an even better experience.

Just like Dronit, Brinder is also 100% open source. It has a beautiful UI and is so complete that you can host it today!

How We built it

We spent day and night in this 3-day period to ensure that what we are delivering is useful, complete and beautiful. We started with building the neural network for Brinder which was indeed the most challenging task. Not only was the concept new but even the research papers that were published on this were either inconclusive or had results that were so far from being accurate that they could solve our purpose and could not fulfill our vision. So we applied all the machine learning techniques we knew and tried at least thrice before we could reach our desired result.

After that we made a beautiful flask app to make our model user accessible, and Parallel to this, we also started working on the geocoding and reverse geocoding technology that we had to use for Dronit. The concept of geocoding straight from a python script was not only difficult but also had been seldom done before. The problem was that it was an essential part of our project. So we made a system that could send and receive Json requests straight from a python script and then used google maps api to decode that request.

Along with this, we used opencv and google's vision api to form a system that could track a dog and click its photograph from a drone's live camera feed.

After we packaged this project, we also made an exquisite web app for it which was dynamic and was updated in real time. At the end, we also made a website to showcase our creations which is live at https://ayush4921.github.io/main.verlan/

Challenges we ran into

Like all great code, ours had a few hiccups in development. After confirming that our proof of concept for brinder would work and after training our initial model in a Google cloud server, we found that our accuracy was only 70%. Clearly, we needed to optimize our code, so we added more layers, switched to the Adam optimizer, used slower learning rates, and trained on a gpu. With these edits, our model was able to reach a 98% test accuracy in literal minutes. However, our validation accuracy was still quite low, but a complete revamp and training for 8+ hrs resulted in over a 90% validation accuracy and a 99% test accuracy which is quite remarkable.

We also had to brainstorm a geocoding system in python from scratch because as far as we know, nobody on the internet yet had made a fool proof open source solution.

Accomplishments that we're proud of:

Making a world class machine learning algorithm having a remarkable, never know before, accuracy.

Making a foolproof open source geocoding system for python.

Presenting our creation with a beautiful UI interface.

Doing all of the above in just 3 days in which us students had to complete our assignments too.

What We learned

Throughout this project our team learned how to pair python code with a good looking UI. We also learned how to make more effective machine learning models using transfer learning and training on cloud servers and how to scrape google images for larger datasets. We also learned about cool technologies such as geocoding and server video streaming.

What's next for Verlan

The most important next step for us will be to access larger datasets to train the model to be able to recognize cats and maybe even other mammals to aid in detecting wildlife presence after natural disasters like wildfires. Having publicly available drones operating will also be useful and attract more people to the site.

How to run Verlan

Just install the requirements for python and run main.py Dronit will be accessible at http://127.0.0.1:5000/adopt and Brinder will be accessible at http://127.0.0.1:5000/ Our main website will be located in index.py The script to run on the drone will be found as drone.py

Built With

Share this project:

Updates