Image-based 3D reconstruction aims to recover three-dimensional scene geometry from visual observations. Traditional reconstruction methods rely on explicit geometric optimization to establish consistency across multiple views, while recent learning-based approaches increasingly use learned geometric priors to reconstruct 3D scenes. A central challenge is how to incorporate geometric constraints into learned representations to achieve accurate and consistent reconstruction while reducing reliance on computationally expensive optimization.

This proposal presents my research on image-based 3D reconstruction through three stages. First, we investigate geometric perception methods for robust depth estimation and feature matching. Second, we develop optimization-based reconstruction approaches that integrate multi-view geometric constraints and learned priors to recover consistent 3D geometry. Third, we explore feed-forward foundation models that incorporate camera projection modeling into learned representations, enabling reconstruction across diverse camera types. These contributions demonstrate how geometric constraints and learned representations can be integrated to advance accurate, consistent, and efficient 3D reconstruction.