DiMViDA: Diffusion-based Multi-View Data Augmentation

We present DiMViDA, a Diffusion-based Multi-View Data Augmentation method built upon an innovative approach for Novel View Synthesis, which uses an extension of diffusion generative models that accepts any number of input views and that can generate any number of missing output views. In this work, our goal is to analyze the benefits of such a generative model in the context of object classification. Given a single input view, we compare the object classification performance of state-of-the-art models, namely ResNet18 and MobileNetV3, using the input view, versus its application to novel views synthesized by our generative model, using such synthetic views to augment the training set. Notably, differently from other works, we also adopt such a multi-view data augmentation method at inference. Our experimental findings illustrate that novel view synthesis can enhance object classification capabilities.