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. We extensively evaluate various aspects of our approach and present its applications for distinctiveness-guided shape retrieval, sampling, and view selection in 3D scenes.
Introduction
In 3D geometry, certain parts of an object are more "meaningful" than others for identification. For instance, the wings and tail of an airplane are more distinctive than its fuselage when distinguishing it from other object classes like cars. Identifying these regions is crucial for tasks such as object retrieval and shape matching.
Traditionally, detecting these regions required either hand-crafted geometric features, which often fail to generalize, or supervised learning, which necessitates massive amounts of human-labeled data. Labeling 3D data is notoriously difficult and time-consuming. This work introduces the first unsupervised deep learning framework to identify distinctive regions on 3D shapes—parts that make an object stand out—by training a network to cluster shapes without any manual labels.
Methodology
The core idea is that for a network to successfully cluster shapes without labels, it must learn to focus on the parts that differentiate one group from another. The proposed pipeline treats distinctive region detection as a byproduct of this unsupervised clustering task.
Feature Embedding and Attention
The architecture uses a modified PointCNN to extract per-point features. These are then processed through a channel-spatial attention unit. This unit allows the network to weigh the importance of different points globally, moving beyond purely local geometric features like curvature. The resulting refined features are fused into a global shape descriptor via average pooling.
Unsupervised Learning Strategy
Since no labels are available, the network employs a clustering-based nonparametric softmax classifier. It iteratively clusters the global features of the shapes in the dataset, treating cluster centers as "pseudo-class prototypes." The network is then trained to classify shapes into these evolving clusters. To stabilize learning, especially in early stages, an adapted contrastive loss is used to pull together different samplings of the same object while pushing away shapes from different clusters.
Extracting Distinctiveness
Once trained, the model's activations reveal the distinctive regions. Points that contribute most significantly to the successful clustering of a shape are deemed the most "distinctive." Notably, distinctiveness is relative; a region is distinctive in relation to the reference dataset. The model successfully adapts to different sets, identifying features that differentiate the current shape from others in the collection.
Key Results and Applications
The method was evaluated on benchmarks like ModelNet40 and the Princeton Shape Benchmark (PSB).
Shape Retrieval
The approach excels at fine-grained, intra-class retrieval. By focusing on distinctive features, it can distinguish between subtle variations within a category.
As shown in the retrieval results, when queried with an airplane, the proposed method successfully finds other airplanes with similar specific features (like engine placement), whereas general unsupervised methods like FoldingNet may only match overall volume.
Distinctiveness-Guided Sampling
To validate the importance of the detected regions, the authors performed downsampling. Keeping only the "distinctive" points allowed for high classification accuracy (~83% with only 256 points), significantly outperforming random or curvature-based sampling. Furthermore, user studies showed that the regions identified by the algorithm (e.g., handles of a cup, strings of a guitar) align closely with regions manually marked by humans.
Conclusion
This unsupervised framework provides a robust way to identify discriminative features in 3D data without the need for costly manual labeling. While the method currently faces some challenges with global symmetry and bridging the gap to human-defined semantic parts, it demonstrates significant potential for improving 3D shape analysis, retrieval, and scene understanding.
Summary updated