Compact Model Representation for 3D Reconstruction

TL;DR

This paper presents a novel method for compactly representing 3D CAD models to enable realistic 3D reconstruction from single images, addressing the challenge of limited information and generalization to unseen objects.

Contribution

It introduces a new approach combining free-form deformation registration and model refinement for efficient 3D mesh representation and reconstruction.

Findings

Achieves dense, realistic 3D reconstructions from single images.

Demonstrates effective compact representation of large CAD model collections.

Provides comprehensive quantitative and qualitative analysis.

Abstract

3D reconstruction from 2D images is a central problem in computer vision. Recent works have been focusing on reconstruction directly from a single image. It is well known however that only one image cannot provide enough information for such a reconstruction. A prior knowledge that has been entertained are 3D CAD models due to its online ubiquity. A fundamental question is how to compactly represent millions of CAD models while allowing generalization to new unseen objects with fine-scaled geometry. We introduce an approach to compactly represent a 3D mesh. Our method first selects a 3D model from a graph structure by using a novel free-form deformation FFD 3D-2D registration, and then the selected 3D model is refined to best fit the image silhouette. We perform a comprehensive quantitative and qualitative analysis that demonstrates impressive dense and realistic 3D reconstruction from single images.