[Paper Review] Unsupervised Detection of Distinctive Regions on 3D Shapes

Xianzhi Li, Lequan Yu+3 more|arXiv (Cornell University)|May 5, 2019

3D Shape Modeling and Analysis60 references4 citations

TL;DR

This paper proposes an unsupervised deep learning method to detect distinctive regions on 3D shapes using point clouds, leveraging a joint clustering-based softmax and contrastive loss to learn per-point distinctiveness without labels. The key contribution is a self-supervised framework that identifies region-level distinctiveness relative to shape classes, enabling applications in retrieval, sampling, and view selection with state-of-the-art performance on ModelNet40.

ABSTRACT

This paper presents a novel approach to learn and detect distinctive regions on 3D shapes. Unlike previous works, which require labeled data, our method is unsupervised. We conduct the analysis on point sets sampled from 3D shapes, then formulate and train a deep neural network for an unsupervised shape clustering task to learn local and global features for distinguishing shapes with respect to a given shape set. To drive the network to learn in an unsupervised manner, we design a clustering-based nonparametric softmax classifier with an iterative re-clustering of shapes, and an adapted contrastive loss for enhancing the feature embedding quality and stabilizing the learning process. By then, we encourage the network to learn the point distinctiveness on the input shapes. We extensively evaluate various aspects of our approach and present its applications for distinctiveness-guided shape retrieval, sampling, and view selection in 3D scenes.

Motivation & Objective

• To develop an unsupervised method for detecting distinctive regions on 3D shapes without relying on hand-crafted features or annotated data.
• To learn per-point and per-shape features that reflect region distinctiveness relative to other shapes in a given set.
• To enable applications such as distinctiveness-guided shape retrieval, point sampling, and optimal view selection in 3D scenes.
• To address the challenge of distinguishing fine-grained shape differences in intra-class settings where global features are insufficient.

Proposed method

• Represent each 3D shape as a point cloud for flexible and lightweight input representation.
• Design a deep neural network trained on an unsupervised clustering task to learn discriminative per-point and per-shape features.
• Formulate a joint loss combining a clustering-based nonparametric softmax and an adapted contrastive loss to stabilize training and enhance feature quality.
• Use the per-point features from the trained network to assign a distinctiveness score to each point, reflecting its uniqueness within the shape set.
• Project distinctiveness scores back to the original mesh to identify distinctive regions on the surface.
• Apply the learned distinctiveness to downstream tasks such as shape retrieval, sampling, and view selection by maximizing distinctiveness in selected views.

Experimental results

Research questions

• RQ1 Can distinctiveness be effectively learned in an unsupervised manner without any labeled data or pre-defined local features?
• RQ2 How does the network’s ability to detect distinctive regions vary between inter-class and intra-class shape sets?
• RQ3 To what extent do distinctive regions contribute to shape classification and retrieval performance?
• RQ4 Can distinctiveness-guided sampling and view selection outperform baseline methods in preserving shape information?

Key findings

• The method achieves state-of-the-art performance in distinctiveness-guided shape retrieval, outperforming general unsupervised feature learning methods like FoldingNet.
• Distinctive regions detected on the same shape vary significantly depending on the reference dataset: inter-class detection emphasizes global contours, while intra-class detection highlights local features like engines on airplanes.
• Sampling points away from distinctive regions has minimal impact on classification accuracy, confirming that these regions are critical for shape discrimination.
• The network can automatically select optimal views in 3D scenes by maximizing the average distinctiveness of visible points, outperforming random or worst-case views.
• The method detects asymmetrical distinctive regions, suggesting potential for improvement by incorporating shape symmetry priors in future work.