3D Modeling from Wide Baseline Range Scans Using Contour Coherence

Ruizhe Wang, Jongmoo Choi, Gérard G. MedioniCVPR 2014 · DOI

Introduction

Standard 3D registration techniques, such as the Iterative Closest Point (ICP) algorithm, are highly effective when scans have significant overlap and a reasonable initial alignment. However, these methods frequently fail in "wide baseline" scenarios—where scans are taken from vastly different angles (e.g., 90 degrees apart). In such cases, the overlap is often less than 40%, and the traditional "closest point" heuristic leads to incorrect correspondences. This paper introduces the concept of "contour coherence" to enable robust 3D modeling from sparse, wide-baseline data.

Methodology: Robust Closest Contour (RCC)

The core of the proposed approach is the Robust Closest Contour (RCC) algorithm. Rather than attempting to match entire surfaces, RCC focuses on apparent contours—the silhouettes where the object's surface disappears from the sensor's view.

The Robust Closest Contour (RCC) pipeline for matching range scans.

The RCC pipeline consists of three primary stages:

  1. Extraction: Identifying contour points by detecting sharp depth discontinuities in the range data.
  2. Pruning: A critical step that removes contour points likely to be occluded from the other camera's perspective, thereby preventing "ghost" matches.
  3. Bijective Matching in 3D: Correspondences are established in 3D space rather than 2D image space to resolve geometric ambiguities. A bijective constraint is enforced, requiring that point A is the closest to point B and vice versa.

Multi-view Registration: M-ICC and MA-ICC

The authors extend RCC into two multi-view frameworks: M-ICC for rigid objects and MA-ICC for articulated objects.

The Multi-view Iterative Closest Contour (M-ICC) algorithm and its application to a Bach head statue.

M-ICC (Multi-view Iterative Closest Contour) optimizes the alignment of all rigid scans simultaneously by minimizing distances across a "view graph." This global approach prevents the cumulative "drift" error common in incremental methods like KinectFusion.

MA-ICC (Multi-view Articulated ICC) handles non-rigid movement, such as human poses, by incorporating Linear Blend Skinning (LBS). It models the object as rigid segments connected by a skeleton, allowing for full-body reconstruction from as few as four wide-baseline views.

Experimental Results

The algorithms were evaluated on synthetic models (Stanford Bunny, Armadillo) and real-world data (Bach head, mannequin, human subjects).

  • Robustness to Rotation: In tests with the Stanford Bunny, while Trimmed-ICP failed beyond a 30° rotation offset, the proposed method remained successful up to a 54° offset.
  • Efficiency and Accuracy: For a mannequin torso, KinectFusion required approximately 1200 frames and resulted in a 5.17mm median error. In contrast, M-ICC achieved a superior median error of 2.12mm using only 4 frames.
  • Articulated Modeling: MA-ICC successfully reconstructed full human models from four orthogonal views, overcoming the wide-baseline challenges that typically defeat standard shape-coherence methods.

Conclusion

By leveraging the information-dense nature of contours and performing matching in 3D space, the proposed M-ICC and MA-ICC algorithms provide a robust solution for wide-baseline 3D modeling. This approach significantly reduces the data requirements for high-quality reconstruction and eliminates the drift associated with incremental tracking. Future work includes automating initial segmentations for articulated models and applying contour coherence to real-time pose estimation.

Summary updated

BibTeX
@inproceedings{wang2014d,
  author = {Ruizhe Wang and Jongmoo Choi and Gérard G. Medioni},
  title = {{3D Modeling from Wide Baseline Range Scans Using Contour Coherence}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2014},
  doi = {10.1109/cvpr.2014.513},
}
Figure