Map Completion Using Generative Models on a Custom Dataset

Abstract

Generative models have been widely used in various image generation tasks, including the completion of partial maps to enhance vision-based exploration and navigation in robotics. While several generative approaches have been proposed for this purpose, to the best of our knowledge, no prior study has explored the use of ImageGPT or Conditional Variational Autoencoder (CVAE) for predicting unseen map regions based on observed ones. Therefore, this study evaluates these two models for the map completion task. We train the models to generate missing map patches around frontiers, with the goal of improving the selection of exploration points as a future research direction. To facilitate this, we adapt the HouseExpo dataset using a custom data generation pipeline, enhancing the efficiency of the overall process. Experimental results demonstrate the effectiveness of these models in indoor exploration environments similar to HouseExpo. The Conditional VAE achieves higher image completion accuracy in terms of all of the score used in this study (IoU, SSIM, F1-score, Accuracy, Precision, Recall, and FID). Additionally, CVAE offers a faster inference time compared to ImageGPT. Furthermore, the proposed pipeline is adaptable to other datasets and generative models, making it a flexible framework for future research.