PCRP: Unsupervised Point Cloud Object Retrieval and Pose Estimation

Pranav Kadam, Qingyang Zhou, Shan Liu, C.-c. Jay Kuo

This paper presents PCRP, a robust framework for identifying 3D objects and determining their orientation without the need for labeled training data. By evolving the "PointHop" family of algorithms, the authors demonstrate that Successive Subspace Learning (SSL) can rival deep learning in efficiency and accuracy for complex 3D tasks.

Introduction

In robotics and scene understanding, two critical tasks are object retrieval (finding a similar object in a database) and pose estimation (determining an object's position and rotation). Most existing solutions suffer from two primary bottlenecks: a heavy dependency on massive labeled datasets and a high sensitivity to alignment. Many retrieval methods assume the query object and database objects are pre-aligned; if an object is rotated or inverted, standard methods often fail.

PCRP addresses these challenges by creating a system that is both unsupervised and rotation-invariant, allowing it to find a "look-alike" object in a gallery and use that match to calculate the query object's pose accurately.

Methodology

The PCRP pipeline: from input object through feature extraction and VLAD aggregation to retrieval and final pose estimation.

The PCRP (Point Cloud Retrieval and Pose estimation) method consists of three main stages:

  1. Enhanced Feature Extraction (FR-PointHop): The authors improve upon the previous R-PointHop method by integrating FPFH (Fast Point Feature Histograms). This handcrafted descriptor is naturally rotation-invariant. These features are processed through a "Saab transform" to capture multi-scale information.
  2. Global Aggregation (VLAD): To convert thousands of individual point features into a single object "fingerprint," the system uses VLAD (Vector of Locally Aggregated Descriptors). Unlike simple max-pooling, VLAD preserves structural information vital for distinguishing between similar categories, such as a sofa versus a chair.
  3. Registration and Symmetry Handling: Once a match is retrieved, the system performs point-to-point registration using SVD and RANSAC. To handle symmetric objects like cars, the algorithm employs symmetry partitioning based on principal components to prevent flipped correspondences.

Key Insights

  • Hybrid Descriptors: By feeding handcrafted FPFH descriptors into an SSL pipeline, the authors achieved superior results compared to using either method in isolation.
  • Efficiency: Because the method uses SSL rather than backpropagation-based deep learning, training takes under 30 minutes, and the model size is approximately 230 KB—orders of magnitude smaller than typical deep learning models (30–70 MB).
  • Rotation Invariance: The local features are invariant to rotation by design, removing the need for the extensive data augmentation (rotating objects during training) required by standard deep learning approaches.

Experimental Results

The framework was evaluated using the ModelNet40 dataset, focusing on challenging categories: airplane, chair, sofa, and car.

Retrieval performance comparison showing PCRP (VLAD) outperforming supervised and traditional methods, especially under arbitrary rotations.
  • Retrieval Accuracy: Under arbitrary rotations, PCRP (using VLAD) achieved a 63.07% Precision@10, significantly outperforming the original PointHop (19.71%) and the supervised PointNet (53.40%).
  • Pose Estimation: The system proved highly effective at category-level estimation. For example, the mean rotation error for airplanes was as low as 2.98 degrees, outperforming traditional methods like ICP and FGR.

Limitations and Future Work

While PCRP excels at isolated objects, its accuracy is tied to the diversity of the gallery set. If the gallery lacks a sufficiently similar object, registration may fail. The authors suggest using Chamfer distance as a threshold to flag and reject unreliable results. Future research will explore scaling this unsupervised logic to complex, cluttered environments where objects may be partially occluded.

Summary updated

BibTeX
@inproceedings{kadam2022pcrp,
  author = {Pranav Kadam and Qingyang Zhou and Shan Liu and C.-C. Jay Kuo},
  title = {{PCRP: Unsupervised Point Cloud Object Retrieval and Pose Estimation}},
  booktitle = {ICIP},
  year = {2022},
  doi = {10.1109/icip46576.2022.9897720},
}
Figure