An application that transforms a simple landscape sketch into a realistic image, which is then mapped into three-dimensional space to provide an estimation of the entire landscape and its environment. My team developed this solution using a custom Conditional GAN (CGAN) model that we trained, alongside pre-trained models for additional functionalities such as depth estimation.

My work

       • Trained a CGAN model for generating the landscape images from sketches
       • Integrated other, pre-trained models into the application for depth-estimation
       • Reconstructed an estimation of the 3D scene from the landscape image

What I learned

       • How to design, train, and evaluate CGAN models, as well as hyperparameter tuning
       • How to work with 3D point clouds
       • How to reconstruct meshes using Poisson Surface reconstruction
‍