-
-
Home Page
-
Our beautiful UI
-
Place where the magic happens
-
About us
-
Landscapes Generated by our model from random color inputs
Inspiration
Our inspiration came from wanting to explore the unknown. In particular, we thought that human imagination was the most unknown space. So we created an AI website that can visually express the landscape that people imagine.
What it does...
ColorScape helps us to explore unknown landscapes using AI through color code data that is received in the form of five colors from the users. It can also function to know various colors of nature. It helps you build landscapes that may not be present on our planet. It lets you peek into your realm of imagination on how other planets or places look like.
How we built it...
We are using conditional generative adversarial networks (cGANs) to accomplish this task. GANs are one of the most effective machine-learning models for new controlled data generation. We used a Dataset with about 4000+ landscape images and the top 5 most abundant colors in those images as input to the GAN. We rescaled and preprocessed the images using OpenCV to a uniform size. We used TensorFlow to train the model. We are using the deep-branched convolutional network.
Dataset: Click Here
Challenges we ran into...
We first got to know each other because of this hackathon, and after creating a group, we found that we had so many different levels of programming skills. Therefore, setting a topic and starting development took a lot of time and discussion. After that, we set up the design of the website and thought about how to get five colors from people. There was a constant error in creating a color picker at the front end and creating a box block to store and display the colors selected by the user. One of the biggest challenges while developing ColorScape was, the amount of data we had was really small. But it was necessary due to time constraints and computational constraints. A bigger network with millions of LandScape images trained for a long time can produce.
Accomplishments that we're proud of...
Team members actively shared, understood, and participated in even areas they did not know in detail. Integration of website development with GANs, which are constantly updated with each other, helped us develop images with better resolutions over time with the selection of colors.
What have we learned?
We each got to learn how to communicate effectively through a virtual platform, manage time effectively and provide teamwork and support from different disciplines and expertise within the 36-hour session. We all took an abstract theme of SwampHack 2023 “Exploring the Unknown” and made it a project utilizing GANs and web design to collaborate on a project to determine what these tools are and what to use them for.
What's next for ColorScope?
As we already talked about with a larger dataset and more training we can produce some amazing unknown landscapes. We can also use this to experience how a landscape will be on an exoplanet. We can train the model with colors, type of terrain, and atmosphere since this information is easy to extract and we can get a really good idea about how mountains, rivers, and other landscapes may look on the planet. GANs are really powerful tools that not only can be used for interactive image generation but for text-to-image generation, enhancement of image resolution, and prediction of the next frame in a video and we hope to dive deeper into it in the future.
** By Chintan Acharya, Seoyoung Kong, and Abelardo D. Montalvo **
Team Leader’s Discord: Chintan Acharya#4058
Log in or sign up for Devpost to join the conversation.