Introduction
Face clustering is a fundamental task in computer vision, essential for organizing large-scale, unlabeled datasets such as those found in surveillance or web-based media. While Graph Convolutional Networks (GCNs) have recently set new benchmarks in this field by leveraging neighborhood relationships, they often introduce significant computational overhead. Traditional GCN-based methods typically rely on complex subgraph construction and non-linear layers, which hinder scalability and speed.
To address these challenges, this paper introduces CSGCN (Confidence-Based Simple Graph Convolutional Network). CSGCN is a streamlined framework that eliminates redundant subgraph processing and simplifies the network architecture to achieve state-of-the-art accuracy with a dramatic increase in processing speed.
The CSGCN Framework
The core philosophy of CSGCN is that "less is more." By removing non-linearities and treating the entire graph as a single entity, the model avoids the "hyperparameter hell" associated with subgraph selection.
1. Linear Graph Convolution
Inspired by Simple Graph Convolution (SGC), the authors remove non-linear activation functions (like ReLU) between layers. This transforms the embedding process into a linear aggregation of neighbor features, significantly reducing training and inference time without sacrificing the discriminative power needed for clustering.
2. Subgraph-Free Processing
Unlike previous methods that break a large graph into thousands of overlapping subgraphs, CSGCN processes the global graph directly. This approach eliminates the computational redundancy of re-calculating features for nodes that appear in multiple subgraphs.
3. Confidence-Based Scoring and Pivot Strategy
To maintain high accuracy despite the simplified architecture, the researchers introduced a Confidence Score ($s$).
- Supervised (CSGCN-S): The score measures how well a node aligns with its neighbors versus nodes of different classes.
- Unsupervised (CSGCN-U): The score is derived from a density-based radius approach.
These scores identify "pivot faces"—high-quality images that serve as reliable cluster centers. A weighted merge-search algorithm then links remaining nodes to these pivots, preventing the "drift" often caused by blurry or occluded images.
Experimental Results
The model was evaluated on major benchmarks including MS-Celeb-1M, YouTube-Faces, and DeepFashion. The results demonstrate a significant leap in efficiency:
- Efficiency: On the MS-Celeb-1M dataset, CSGCN-S completed the clustering task in just 3.8 minutes. In contrast, previous state-of-the-art methods like L-GCN and GCN-D required 86.8 and 62.2 minutes, respectively—making CSGCN up to 20x faster.
- Accuracy: CSGCN-S achieved a Bcubed F-score of 87.01 on MS-Celeb-1M, outperforming more complex GCN variants.
- Unsupervised Strength: In unsupervised settings, CSGCN-U reached an F-score of 90.54 on YouTube-Faces, far exceeding traditional algorithms like Approximate Rank-Order (ARO), which scored 75.42.
Conclusion
CSGCN demonstrates that for face clustering, high-dimensional non-linearity is often unnecessary if a robust confidence mechanism is in place. By simplifying the GCN architecture and focusing on high-confidence "pivot" nodes, the framework provides a highly scalable and accurate solution for real-world, large-scale face organization tasks. Future work will focus on enhancing robustness against extreme image noise and further refining feature extraction from unlabeled data.
Summary updated