One Shot Learning for Edge Detection on Point Clouds

Zhikun Tu, Yuhe Zhang, Yiou Jia, Kang Li, Daniel Cohen-Or

TL;DR

This paper introduces OSFENet, a novel one-shot learning approach for edge detection on 3D point clouds. By learning the specific data distribution of a target point cloud from a single example, OSFENet achieves state-of-the-art performance, outperforming general-purpose models, especially on real-world scanned data.

What Problem Does This Paper Address?

Edge detection on 3D point clouds is crucial for understanding object shapes and structures. However, existing deep learning methods often struggle due to variations in data acquisition from different scanners, leading to diverse sampling error distributions. Training a single model on a mixed-scanner dataset results in suboptimal performance for any specific scanner. This paper tackles this challenge by proposing a one-shot learning method that adapts to the specific characteristics of the target point cloud data, aiming for superior edge detection accuracy with limited training data.

Methods

The core of the proposed method is OSFENet, a lightweight network designed for one-shot edge detection. It employs a novel filtered-kNN-based surface patch representation to capture local geometric information. Instead of using a large, fixed number of neighbors, it refines the kNN selection to avoid points from opposite sides of thin surfaces, creating more robust local patches.

A key component is the RBF_DoS (Radial Basis Function-based Descriptor of the Surface patch) module. This module integrates Radial Basis Functions (RBFs) to describe the surface patch around each point. It utilizes both Euclidean and cosine distances between neighboring points, processed through RBF operators and fully connected layers, to generate rich surface descriptors. These descriptors, along with the point patch information, are then fed into a 4-layer Transformer encoder for feature extraction. Finally, a decoder predicts the probability of each point being an edge.

Datasets

The paper evaluates OSFENet on several datasets:

  • Object-level Datasets:
    • ABC dataset [8]: Used for performance comparison with baselines.
  • SHREC dataset [34]: Used to demonstrate limitations.
  • Real-scanned Scene Datasets:
    • S3DIS dataset [35]: Indoor scenes.
  • Semantic3D dataset [5]: Large-scale outdoor scenes.
  • UrbanBIS dataset [4]: Large-scale urban scenes.

The authors also mention augmenting training data by rotating it six times along the X, Y, and Z axes, both clockwise and counterclockwise by 90 degrees.

Key Insights and Ideas

  • One-Shot Learning for Point Cloud Edges: The core idea is that learning the specific data distribution of a target point cloud from a single example (or a small segment) leads to significantly better edge detection than training on a general, mixed-scanner dataset.
  • Filtered-kNN for Robust Patches: Standard kNN can be problematic for thin structures. The proposed filtered-kNN approach improves the quality of local surface patches by selecting more relevant neighbors.
  • RBF_DoS for Rich Surface Descriptors: Integrating RBFs with both Euclidean and cosine distance metrics provides a powerful way to capture local surface geometry, which is crucial for distinguishing edges.
  • Transformer for Contextual Feature Encoding: The Transformer architecture effectively processes the local features generated by RBF_DoS, capturing relationships between neighboring points.

Key Results

  • State-of-the-Art Performance: OSFENet outperforms several strong baselines (BE, SGLBP, EC-Net, PIE-Net, NerVE, DEF, NEF) on the ABC dataset (both ABC-NEF and ABC-ALL) across metrics like CD, IoU, Precision, Recall, and F-score. For example, on ABC-ALL, OSFENet achieves an F-score of 0.8997, significantly better than most competitors.
  • Superiority on Real-Scanned Data: The method shows remarkable performance on real-scanned datasets (S3DIS, Semantic3D, UrbanBIS), especially when trained on segments of the target scene, highlighting the advantage of its one-shot learning approach.
  • Robustness: OSFENet demonstrates robustness to Gaussian noise and downsampling, maintaining competitive performance.
  • Efficiency: While not the fastest for very small models, OSFENet offers competitive throughput, especially for larger models where some competitors fail.
  • Qualitative Improvements: Visual results show OSFENet detecting finer and smoother edges compared to baselines, which often miss subtle details or incorrectly identify non-edges.

Limitations and Future Work

  • Dataset Variability: The paper acknowledges that performance might still be affected when dealing with datasets acquired through vastly different methods.
  • Surface Segmentation: While the edge detection is strong, the simple flood-fill approach for surface segmentation can fail in cases of small gaps in detected edges.
  • Future Work: The authors suggest extending RBF_DoS to encompass the entire shape of point clouds for tasks like classification and segmentation. They also propose integrating local and large-scale patches to improve edge detection in cases where local patches fail to capture overall edge shapes.

Summary updated

BibTeX
@article{tu2025one,
  author = {Zhikun Tu and Yuhe Zhang and Yiou Jia and Kang Li and Daniel Cohen-Or},
  title = {{One Shot Learning for Edge Detection on Point Clouds}},
  journal = {IEEE Trans. Vis. Comput. Graph.},
  year = {2025},
  doi = {10.1109/tvcg.2025.3542475},
}

Loading related content...

Figure