Adapt-Infomap: Face Clustering with Adaptive Graph Refinement in Infomap

Abstract

Face clustering is an essential task in computer vision due to the explosion of related applications such as augmented reality or photo album management. The main challenge of this task lies in the imperfectness of image feature representations. Given image features extracted from an existing pre-trained representation model, it is still an unresolved problem that how can the inherent characteristics of similarities of unlabelled images be leveraged to improve the clustering performance. In order to solve face clustering in an unsupervised manner, we develop an effective framework named as Adapt-InfoMap. Specifically, we first reformulate face clustering as a process of non-overlapping community detection. Then, Adapt-InfoMap achieves face clustering by minimizing the entropy of information flows (as known as the map equation) on an affinity graph of images.  Since the affinity graph of images might contain noisy edges, we develop an outlier detection strategy in Adapt-InfoMap to adaptively refine the affinity graph. Experiments with ablation studies demonstrate that Adapt-InfoMap significantly outperforms existing methods and achieves new state-of-the-arts on three popular large-scale datasets for face clustering, e.g., an absolute improvement of more than 10% and 3% comparing with prior unsupervised and supervised methods respectively in terms of average of Pairwise F-score.

Keywords: Face Clustering, Map Equation, Graph Partitioning