Abstract
As 3D technologies advance rapidly, generating a large number of 3D models effectively and efficiently has become a remarkable challenge. This paper proposes a novel editing method based on the feature lines of images (contour and principal axis) for generating new 3D models. By comparing the feature lines of a target image and an original 3D model, constraint conditions are established to control the editing process. The method utilizes as-rigid-as-possible (ARAP) mesh deformation guided by a new energy function to produce models that resemble the target image while maintaining structural integrity. Experimental results demonstrate that this method is more effective and efficient than existing techniques.
Introduction
Creating high-quality 3D models from scratch is a labor-intensive process requiring specialized skills. While "model reuse"—modifying an existing 3D model—is more efficient, current image-based editing methods often struggle with geometric distortion and high computational complexity. Existing algorithms frequently focus only on distances between points, leading to "collapsed" features or unnatural stretching (e.g., a bear’s nose flattening or a cup’s rim curling). The authors aim to bridge the gap between a designer’s 2D vision and a 3D asset, ensuring the final model looks like the target image while maintaining its structural integrity.
Methodology
The researchers built their method on the foundation of As-Rig-As-Possible (ARAP) mesh deformation, introducing two critical innovations:
Similar Triangle Correspondence
Instead of just mapping points based on their position along a contour (length proportion), the system uses the principal axis of the object to form triangles. By ensuring the triangles in the 3D projection are similar to those in the 2D sketch, the algorithm captures angular features. This prevents the "shearing" or "squashing" effects common in older methods.
Global Energy Function ($E_{IARAP}$)
Standard ARAP focuses on local rigidity—keeping small neighborhoods of vertices consistent. However, this can lead to global distortion. The authors added a global energy term ($E_c$) that measures the stability of the entire region's structure. A balancing coefficient ($\gamma$) allows the system to weigh local details against global shape stability. Through empirical testing, a $\gamma$ value of 0.15 was found to provide the best balance.
The Pipeline
- Extract: Get the contour and principal axis from a 2D image/sketch.
- Project: Project the original 3D model onto a 2D plane from a user-selected viewpoint.
- Match: Establish vertex-to-pixel correspondence using the similar triangle rule.
- Deform: Solve a sparse linear equation set to minimize the combined energy function, resulting in the new 3D shape.
Results and Discussion
The method was evaluated using the SHREC12 and Princeton University datasets.
Accuracy and Fidelity
The researchers used a Laplacian distance metric ($E_a$) to measure similarity. Their method consistently achieved lower $E_a$ scores (0.015 to 0.12) compared to previous methods (0.08 to 0.24), indicating a much closer match to the target image.
Computational Efficiency
The proposed method is roughly 2x to 4x faster than existing literature. For example, editing an airplane model took 6.42 seconds, compared to 20.35 seconds and 15.46 seconds for competing methods. This efficiency is achieved by comparing contours only once to establish control points, rather than using iterative cross-section comparisons.
Conclusion
This paper presents an automated image-based approach that allows users to generate high-quality 3D models without manual "point-pulling." While the method currently requires manual viewpoint selection, future work will focus on automating view selection and integrating deep learning to address depth ambiguity in 2D sketches. Despite these limitations, the method provides a significant advancement in making 3D model editing more accessible and efficient.
Summary updated