Quality-driven poisson-guided autoscanning

Shihao Wu, Wei Sun, Pinxin Long, Hui Huang, Daniel Cohen-Or, Minglun Gong, Oliver Deussen, Baoquan Chen

TL;DR

This paper introduces a novel "quality-driven" approach to autonomous 3D scanning that prioritizes high-fidelity reconstruction over simply minimizing scan count. By analyzing a Poisson field's confidence map, the system intelligently selects "Next-Best-Views" (NBVs) to progressively capture intricate geometric details, ensuring both completeness and accuracy in the final 3D model.

What Problem Does This Paper Address?

Acquiring complete and high-fidelity 3D models of physical objects is a challenging task, especially when the object's shape is unknown beforehand. Traditional scanning methods often focus on minimizing the number of scans required to cover the object's surface. However, this can lead to incomplete reconstructions or a lack of detail in complex areas. The core problem this paper addresses is how to autonomously plan a scanning strategy that guarantees a high-quality, detailed 3D model, even for objects with intricate geometry and occlusions, by intelligently deciding where to position the scanner next.

Methods

The paper proposes a "quality-driven Poisson-guided autoscanning" method. The key innovation lies in its focus on reconstruction quality.

  1. Poisson Surface Reconstruction and Confidence Map: The system first reconstructs a tentative, watertight surface model from an initial point cloud using Screened Poisson surface reconstruction. Crucially, it then generates a "confidence map" for this reconstructed surface. This map quantifies the quality and fidelity of the estimated iso-surface at each point, considering factors like the stability of the local topology and the support from raw scan data.
  2. Viewing Vector Field (VVF) Generation: Based on the confidence map, a 3D viewing vector field (VVF) is generated. The magnitude of vectors in this field indicates how suitable a location is for scanning, and their direction suggests an optimal scan orientation.
  3. Next-Best-View (NBV) Selection: The VVF is used to identify "Next-Best-Views" (NBVs). Instead of just looking for areas with missing data, the system prioritizes locations that offer the best potential to improve both the completeness and local smoothness of the reconstructed surface. A greedy algorithm selects a sparse set of NBVs to avoid redundant movements.
  4. Quality-Driven Iteration: The scanning process is iterative. After acquiring data from the selected NBVs, the point cloud is updated, the Poisson surface and confidence map are re-evaluated, and new NBVs are computed. This loop continues until a desired level of quality is achieved or no further improvement is possible.
  5. Selective Point Insertion: To further refine the model and ensure high density in detailed areas, a selective point insertion strategy is employed. New points are added to the point cloud with a probability inversely proportional to the confidence of the nearest iso-point, ensuring that areas requiring more detail receive denser sampling.

Datasets

The paper does not introduce new datasets. Instead, it evaluates its method on real-world physical objects and also uses digital models for simulation and comparison. Specific examples of objects scanned include action figures, animal models, a pirate boat, and a church model.

Key Insights and Ideas

  • Quality over Quantity: The fundamental insight is that prioritizing reconstruction quality (completeness and fidelity) over minimizing scan count leads to significantly better 3D models, especially for complex objects.
  • Poisson Field as a Quality Indicator: Leveraging the properties of the Poisson field and its gradients to derive a confidence map is a powerful way to assess surface quality, going beyond simple visibility or boundary detection.
  • Integrated Completeness and Smoothness: The confidence map effectively combines measures of both surface completeness and local smoothness, providing a holistic view of reconstruction quality.
  • Intelligent NBV Selection: The VVF and NBV selection strategy intelligently guides the scanner to areas that are not only incomplete but also require finer geometric detail, leading to more efficient and effective scanning.
  • Adaptive Sampling: The selective point insertion strategy ensures that detailed areas are sampled more densely, resulting in richer reconstructions without unnecessary data in smooth regions.

Key Results

  • The proposed method successfully generated complete and high-fidelity 3D models of complex physical objects, outperforming state-of-the-art visibility-based and boundary-based NBV approaches in terms of detail and accuracy.
  • Quantitative evaluations showed that the method achieved a higher percentage of points with low reconstruction errors compared to competing approaches.
  • The system demonstrated its effectiveness on two different robotic platforms (PR2 mobile robot and a one-arm industry robot).
  • The autonomous scanning process was significantly faster and required less user effort than manual scanning for comparable results. For instance, an experienced user took over 200 scans and three hours for a pirate boat model, while the autoscanner achieved a comparable result with about 100 scans in less than two hours.
  • The method showed robustness to simulated noise, producing satisfactory results even under significant noise corruption.

Limitations and Future Work

  • Scanner Accessibility: The physical limitations of scanner accessibility (e.g., size and reach) can prevent the scanner from reaching certain areas, even if the algorithm identifies them as needing more data. This can lead to imperfect depth acquisition in such areas.
  • Isotropic Sensor Assumption: The current approach assumes an isotropic sensor, meaning it doesn't explicitly account for the sensor's orientation capabilities beyond basic direction. This can limit performance in scenarios where specific orientations are crucial.
  • Greedy Approach: The current NBV selection is greedy, aiming to improve the lowest confidence areas at each step. The authors suggest that more efficient scanning sequences might exist to reach the same quality faster or with fewer scans.
  • Future Work:
    • Explore searching in a 6D pose space (including rotation) for more optimal scanning.
  • Investigate incorporating more advanced sensor models.
  • Develop semi-automatic interactive scanning mechanisms where the confidence map can serve as visual feedback for human operators.

Summary updated

Abstract

We present a quality-driven, Poisson-guided autonomous scanning method. Unlike previous scan planning techniques, we do not aim to minimize the number of scans needed to cover the object's surface, but rather to ensure the high quality scanning of the model. This goal is achieved by placing the scanner at strategically selected Next-Best-Views (NBVs) to ensure progressively capturing the geometric details of the object, until both completeness and high fidelity are reached. The technique is based on the analysis of a Poisson field and its geometric relation with an input scan. We generate a confidence map that reflects the quality/fidelity of the estimated Poisson iso-surface. The confidence map guides the generation of a viewing vector field, which is then used for computing a set of NBVs. We applied the algorithm on two different robotic platforms, a PR2 mobile robot and a one-arm industry robot. We demonstrated the advantages of our method through a number of autonomous high quality scannings of complex physical objects, as well as performance comparisons against state-of-the-art methods.
BibTeX
@article{wu2014quality,
  author = {Shihao Wu and Wei Sun and Pinxin Long and Hui Huang and Daniel Cohen-Or and Minglun Gong and Oliver Deussen and Baoquan Chen},
  title = {{Quality-driven poisson-guided autoscanning}},
  journal = {ACM Trans. Graphics},
  year = {2014},
  url = {https://openreview.net/forum?id=WAniIHiXfo-W},
}
Figure