Awards & Nominations
Nepali Ho Ni has received the following awards and nominations. Way to go!
Global Nominee
jov.ai
High-Level Project Summary
The grayish tone and B/W images, seizing the ravishing views from the jovian system including the unique discoveries and snapshots of the various elements in the circumference, captured by Junocam could seem plain at times. These processed portraits are an immense source of discoveries, educational approaches, and awareness. So, keeping all these in our mind, we built an image editing application that primarily performs 2 main functions. 1. convert b/w images to colored images highlighting the details of Jupiter's surface, while giving it an artistic touch and 2. generate new b/w images like the ones Junocam captures.
Link to Final Project
Link to Project "Demo"
Detailed Project Description
jov.ai is an AI tool that helps to enhance Black and White images captured by Junocam. It does so by using Deep Learning tools. We also truncated StyleGAN to make a microStyleGAN which generates new images that resemble the snapshots taken by Junocam.
The project aims to bring artistic enhancements to the pictures taken by Junocam and help create more alike pictures as a reference to researchers of all potential conditions.
We used Python programming language, PyTorch, TesorFlow, Numpy, and other libraries to create our deep learning models. The microStyleGAN was made in PyTorch and the color enhancement was a previous project(in references) built in tensorflow.
As the models were computationally heavy, we used Kaggle kernel to train the microStyleGAN and we used pretrained weights for the color enhancement. We used free hosting to show our work, due to which we cannot upload the models to have maximum user interaction. However, the website does show the entire workflow and would give a strong feel on how the project works.
Space Agency Data
We have used images of Jupiter from the Juno spacecraft for testing our model. We used the pretrained colorization model and used images from the Junocam camera to test it.
The microStyleGAN was trained based on 8000 raw, black and white images at kaggle repository. This dataset contains different images taken under Juno Mission by NASA. It helped us to easily gather a lot of images to train the model which was a huge time savior for us, otherwise we would have to write a script to manually download recent images, and then upload it to somewhere online again.
Hackathon Journey
This space Apps Challenge and our hackathon journey together was a rollercoaster full of ups, and downs along with twirls. There is no space for doubt that we undeniably learned and excelled a bit more in our specific terrains whether it is About coding, designing, calculations, or training. But we presume our jaunt was much more than technical things. What impressed us at the end of the day was time management. Yes, scarcity indeed startled us, but proper allotment of responsibilities and communication was worth it. Reminiscing the inception of this teamwork, all of us were people from different places and fields to this date, but we now introduce one another as a team, team “Nepali Ho ni”. This challenge made us apprehend the mistakes, redos, downfalls, and the rise after it. The most crucial virtue we learned was hard work and honesty. Even though there were times when we were on the verge of falling apart in conquering the deep learning models, we understood that everyone giving their best was way above all those sweats and headaches. Summing up everything, this space app challenge and its journey was an overwhelming voyage for everyone.
References
http://richzhang.github.io/colorization/
Colorful Image Colorization arXiv:1603.08511
A Style-Based Generator Architecture for Generative Adversarial Networks, arXiv:1812.04948
MicrostyleGAN from GAN Specialization at coursera(course 2: build better GAN)