Align then clip: Refining graph for face clustering.

Abstract

Face clustering, a critical task for annotating large-scale unlabeled face recognition datasets, aims to group facial images of the same identity while minimizing annotation costs. Recent approaches model face relationships as graphs and leverage graph neural networks (GNNs) to capture structural dependencies. However, the performance of these methods heavily relies on the quality of the input graph, where conventional kNN-based graph construction suffers from two key limitations: 1) sensitivity to hyperparameter k, leading to either excessive false-positive edges (large k) or fragmented true-positive connections (small k), and 2) propagation of noise through message passing in GNNs. To address these challenges, we propose AtC (Align then Clip), a novel framework that refines graph structures through dual-phase optimization. During training, we introduce a distribution alignment branch that aligns node representations between noisy and clean graphs, enhancing robustness to edge noise. For inference, we design Post-Clipping, an adaptive edge pruning strategy that eliminates spurious connections while preserving critical linkages. Furthermore, our Post-Linkage mechanism mitigates cluster fragmentation by reconnecting split components. Extensive experiments validate its scalability (e.g., 5.21M samples), robustness to 30 % false-positive noise, and superior cross-domain generalization, demonstrating broad applicability in real-world scenarios. Extensive experiments demonstrate the method's scalability (on datasets of up to 5.21 million samples), robustness (even with 30 % false-positive noise), and superior cross-domain generalization. These results showcase its strong potential for real-world applications.