Unsupervised Feature Learning With Graph Embedding for View-Based 3D Model Retrieval

Yuting Su, Wenhui Li, Weizhi Nie, Dan Song, An-An Liu

Abstract

Efficient 3D model retrieval is critical for applications in virtual reality, computer-aided design, and gaming. However, current methods often struggle with the high cost of manual labeling or the computational inefficiency of complex graph-matching algorithms. This paper presents the Unsupervised Feature Learning with Graph Embedding (FLGE) framework. FLGE iteratively refines model relationships through graph embedding, learning discriminative features without human supervision. By projecting similarity measures directly into the feature space, FLGE achieves state-of-the-art accuracy while delivering a 126x speedup over traditional graph-based methods, making it highly suitable for real-time retrieval in large-scale databases.

Introduction

The explosion of 3D data has made view-based 3D model retrieval—identifying objects based on multiple 2D projections—a primary research focus. Despite progress, two major challenges remain: the "labeling bottleneck," where deep learning models require vast amounts of annotated data, and the "speed-accuracy trade-off," where unsupervised methods must choose between fast but simplistic distance metrics or accurate but slow graph-matching techniques.

The proposed FLGE framework addresses these issues by learning features in an unsupervised manner and embedding similarity calculations into the feature space itself, ensuring both high precision and real-time performance.

The FLGE Framework

The FLGE framework operates through an iterative loop consisting of three primary modules: graph refinement, label estimation, and metric learning.

1. Graph Construction and Refinement

Initially, 3D models are represented using features from a pre-trained AlexNet. To improve the initial noisy graph, the system employs k-reciprocal nearest neighbors. This technique assumes that if Model A is a top neighbor of Model B and vice versa, they likely belong to the same category. This filtering process effectively removes visual outliers and strengthens the underlying structure of the data.

Graph Refinement: Iteratively updating relationships between models to filter noise and strengthen category clusters.

2. Label Estimation

Without ground-truth labels, FLGE estimates "pseudo-labels" to guide the learning process.

  • Positive Labels: A Gaussian kernel assigns higher weights to models that are proximal within the refined graph.
  • Hard Negative Mining: Instead of random selection, the system identifies "hard negatives"—models that appear dissimilar to the query but share similarities with the query's neighbors—to challenge and improve the metric learning process.

3. Metric Learning

The system learns a projection matrix that transforms the original feature space into a new embedding where similar models are clustered and dissimilar ones are separated. This phase essentially "bakes" the complex graph relationships into the feature vectors themselves.

Metric Learning: Projecting features into a space where similar objects are pulled together and dissimilar ones are pushed apart.

Experimental Evaluation

The researchers evaluated FLGE on three standard benchmarks: NTU (549 models), MV-RED (505 objects), and ETH (80 objects).

Key Results

  • Superior Accuracy: FLGE consistently outperformed baseline methods. On the MV-RED dataset, it achieved a 15.1% improvement in First Tier (FT) recall.
  • Massive Speedup: On the NTU dataset, FLGE reduced query time from 31.63 seconds (using the previous HGS method) to just 0.25 seconds. This 126x speedup demonstrates the framework's practical utility for real-time applications.
  • Feature Clustering: Visualizations confirm that the iterative process successfully transforms disorganized initial features into distinct, category-specific clusters.

Discussion and Conclusion

The core innovation of FLGE lies in treating similarity as a learnable feature rather than a post-processing step. By utilizing contextual information through k-reciprocal neighbors and robust hard negative mining, the system overcomes the limitations of raw visual features.

While the framework is sensitive to initial feature quality and hyper-parameter selection (specifically the number of neighbors $k$), its performance and speed represent a significant leap forward. Future work may involve extending this iterative graph-embedding approach to cross-modal retrieval tasks, such as retrieving 3D models from text descriptions.

Summary updated

Abstract

3D model retrieval is becoming a hot research topic due to its wide applications such as computer-aided design, digital entertainment, and virtual reality. For this challenging task, feature learning and similarity measure are two critical problems. However, existing approaches usually learn discriminative visual features and develop a complex graph matching strategy to measure the similarity independently. In this paper, we propose an unsupervised method which can embed similarity measure into the feature space. The proposed method utilizes both similarity and dissimilarity information to better leverage the unsupervised problem and estimates the labels which are further used for metric learning. With the learned metric, we project the original features to more discriminative feature space and efficiently measure the similarity among models under the new feature space. We conduct extensive evaluations of three popular and challenging datasets. The experimental results demonstrate the superiority and effectiveness of the proposed method, competing against the state of the arts.
BibTeX
@article{su2019unsupervised,
  author = {Yuting Su and Wenhui Li and Weizhi Nie and Dan Song and An-An Liu},
  title = {{Unsupervised Feature Learning With Graph Embedding for View-Based 3D Model Retrieval}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2929109},
}
Figure