Towards Diffusion Model Based Dataset Augmentation for Negative Obstacle Detection Systems

This work presents a methodology to generate synthetic images of negative obstacles such as potholes to train artificial vision systems for autonomous vehicles. The proposed methodology uses Latent Diffusion Models and Dreambooth LoRAs to fine-tune models to includer subjects of which we have few photographs. The method is validated by training a simple predictive model with a subset of the synthetically generated images and testing it with a mixed training dataset of real and synthetic images. The model shows good precision in distinguishing roads with and without potholes. As such, the proposed methodology demonstrates the plausibility of generating realistic images with the objective to train artificial vision models focused in autonomous vehicles whit some limitations such as the generation of unusable images due to aberrations and the labelling requirements.