CD ² : Fine-grained 3D Mesh Reconstruction with Twice Chamfer Distance

Monocular 3D reconstruction is to reconstruct the shape of object and its other information from a single RGB image. In reconstruction, polygon mesh, with detailed surface low computational cost, most prevalent expression form obtained deep learning models. However, state-of-the-art schemes fail directly generate well-structured meshes, we identify that meshes have severe Vertices Clustering (VC) Illegal Twist (IT) problems. By analyzing mesh deformation process, pinpoint inappropriate usage Chamfer Distance (CD) loss root cause VC IT problems in model. this article, initially demonstrate these two induced by CD visual examples quantitative analyses. Then, propose fine-grained method 2 employing distance twice perform plausible adaptive deformation. Extensive experiments on datasets comparisons five latest our generates outperforms others terms several metrics.