Is Deep Diffusion Probabilistic Model Applicable for Fingerprint-based Indoor Localization?

Abstract

The latest deep learning (DL) phenomenon is the Denoising Diffusion Model (DDM). DDM is in a class of latent variable models of the deep generative model (DGM) along with the big name of Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs). Moreover, in a recent finding, DDM beats GANs in image synthesis. This paper presents the prospective applicability discussion of DDM for indoor localization research as previous models, e.g., GANs and VAEs, which are successfully implemented. Here, we focus more on how DDM can synthesize localization parameters with the help of fingerprinting technique's database enhancement. The fingerprint technique needs a preconstructed database which has the main drawbacks of its cost, time inefficient, and high complexity. We found valuable works of literature on this specific topic for GANs and VAEs. However, there are few DDM applications for discrete data types, and as the authors' concern, there is no attempt to apply them to indoor localization yet. DDM implementation is to generate continuous data domains, e.g., image, text, and audio data. A radio map or fingerprint database is essentially needed for fingerprint-based indoor localization. Learning this database pattern helps increase the system's performance. Obtaining a high-density and quality database is expensive and challenging to implement. Then, it raises a question, is DDM applicable for synthesizing this database and alleviating this problem?