Complex 3D shape recovery using hybrid geometric shape features in a hierarchical shape segmentation approach

Hongwei Zheng, Dietmar SaupeWorkshops @ ICCV 2009 · DOI

Reconstructing complex 3D objects from high-resolution scans often leads to global alignment errors (drift) and memory bottlenecks. This paper, originally presented at ICCV Workshops in 2009, proposes a hierarchical segmentation approach that utilizes a low-resolution global scan as a geometric prior. By segmenting the global model into simpler geometric primitives, high-resolution local patches can be accurately registered and seamlessly integrated, overcoming traditional limitations in 3D shape recovery.

The Challenge of 3D "Big Data"

The primary hurdle in 3D scanning is balancing high-resolution detail with global structural integrity. Traditional methods that align local patches sequentially often suffer from accumulated errors, leading to "drift." Furthermore, the massive data volume of high-resolution scans—often exceeding 1GB—frequently exceeds available memory for global optimization. This paper introduces a "best of both worlds" strategy using hybrid data and hierarchical segmentation.

Methodology: A Divide-and-Conquer Approach

The proposed pipeline relies on two core innovations: hybrid data acquisition and a hierarchical segmentation strategy.

Hybrid Data Acquisition

Researchers utilize two distinct scanning technologies to capture the object. A stationary scanner (Minolta VIVID-900) captures high-resolution local surface patches, while a hand-held laser scanner (Polhemus FastSCAN) provides a rapid, low-resolution "global" model of the entire object. This low-resolution model serves as a "geometric prior" or anchor for the high-resolution data.

Hierarchical Shape Segmentation

To manage memory and improve robustness, the global model is segmented into meaningful sub-shapes. The system fits geometric primitives—planes, spheres, and cylinders—to the mesh. This allows the object to be processed part-by-part (e.g., a leg, then the tail), significantly reducing the computational load at any single step.

The proposed pipeline: from hybrid data acquisition and hierarchical segmentation to the final high-resolution reconstruction of an anteater skeleton.

Registration Pipeline

  1. Rigid ICP: Local high-resolution patches are first aligned to the global low-resolution model using the Iterative Closest Point (ICP) algorithm.
  2. Non-rigid TPS Adjustment: Thin-Plate-Splines (TPS) are used to warp the high-resolution patches slightly, ensuring they conform perfectly to the global geometric prior. This step diffuses local alignment errors and creates a seamless surface even if individual scans have slight distortions.

Case Study: The Anteater Skeleton

The system was demonstrated on an anteater skeleton approximately 0.80 meters long. The raw data consisted of 370 high-resolution scans totaling 1.162 GB (30 million vertices). Through the hierarchical approach, the final model was reduced to a manageable 293 MB with 5.2 million vertices while preserving critical anatomical details.

Key Results and Performance

The hybrid method significantly outperforms standard pairwise registration. By using the global scan as an anchor, the "drift" common in traditional methods is eliminated. Qualitatively, the approach produces complete, detailed models without the gaps or "holes" often seen when high-resolution patches are registered in isolation.

Qualitative comparison of detail recovery in a complex joint structure, showing the transition from raw data to a refined, segmented reconstruction.

Memory Efficiency and Robustness

The hierarchical approach allowed the researchers to process over 1GB of raw data—a significant feat for the time—by never needing to hold the entire high-resolution dataset in memory at once. The system successfully handled 370 separate scans, automatically correcting large alignment errors that would typically require extensive manual intervention.

Conclusion

By combining low-resolution global priors with high-resolution local patches through hierarchical segmentation, this approach provides a robust solution for complex 3D reconstruction. While currently dependent on geometric primitives and requiring some manual initialization, it represents a significant step forward in handling large-scale, high-detail 3D scanning data.

Summary updated

BibTeX
@inproceedings{zheng2009complex,
  author = {Hongwei Zheng and Dietmar Saupe},
  title = {{Complex 3D shape recovery using hybrid geometric shape features in a hierarchical shape segmentation approach}},
  booktitle = {Workshops @ ICCV},
  year = {2009},
  doi = {10.1109/iccvw.2009.5457483},
}
Figure