Dynamic Grouping With Multi-Manifold Attention for Multi-View 3D Object Reconstruction

In a multi-view 3D reconstruction problem, the task is to infer the 3D shape of an object from various images taken from different viewpoints. Transformer-based networks have demonstrated their ability to achieve high performance in such problems, but they face challenges in identifying the optimal way to merge the different views in order to estimate with great fidelity the 3D shape of the object. This work aims to address this issue by proposing a novel approach to compute information-rich inter-view features by combining image tokens with similar distinctive characteristics among the different views dynamically. This is achieved by leveraging the self-attention mechanism of a Transformer, enhanced with a multi-manifold attention module, to estimate the importance of image tokens on-the-fly and re-arrange them among the different views in a way that improves the viewpoint merging procedure and the 3D reconstruction results. Experiments on ShapeNet and Pix3D validate the ability of the proposed method to achieve state-of-the-art performance in both multi-view and single-view 3D object reconstruction.