VICTOR ESCRIBANO GARCIA'S PORTFOLIO
About the project
I developed a system called IPM_GAN, which generates top-view maps (Inverse Perspective Mapping) from front camera views using Generative Adversarial Networks (GANs). This project leverages the power of GANs to transform front-view images of roads into top-view maps, facilitating enhanced understanding and navigation.
The dataset for this project was created using raw images from the KITTI Dataset. To achieve accurate top-view transformations, I implemented a custom calibration tool to extract the homography matrix. This tool allows users to calibrate the system by selecting corresponding points on source and target images, ensuring precise mapping for dataset generation. After filtering out consecutive frames to avoid repetition, the tool generates the necessary target images for training the model.
Training the IPM_GAN model involves using the curated dataset to develop a robust transformation system. The results demonstrate the model's effectiveness in generating accurate top-view maps from front-view images, as shown in various test cases. The model's performance is further illustrated by the generator and discriminator loss plots, as well as the discriminator accuracy metrics, highlighting its success in achieving the project's objectives.