MGH: Metadata Guided Hypergraph Modeling for Unsupervised Person Re-identification

Abstract

As a challenging task, unsupervised person ReID aims to match the same identity with query images which does not require any labeled information. In general, most existing approaches focus on the visual cues only, leaving potentially valuable auxiliary metadata information (e.g., spatio-temporal context) unexplored. In the real world, such metadata is normally available alongside captured images, and thus plays an important role in separating several hard ReID matches. With this motivation in mind, we propose MGH, a novel unsupervised person ReID approach that uses meta information to construct a hypergraph for feature learning and label refinement. In principle, the hypergraph is composed of cameratopology-aware hyperedges, which can model the heterogeneous data correlations across cameras. Taking advantage of label propagation on the hypergraph, the proposed approach is able to effectively refine the ReID results, such as correcting the wrong labels or smoothing the noisy labels. Given the refined results, We further present a memory-based listwise loss to directly optimize the average precision in an approximate manner. Extensive experiments on three benchmarks demonstrate the effectiveness of the proposed approach against the state-of-the-art. CCS CONCEPTS• Information systems → Image search.

Offline-Online Associated Camera-Aware Proxies for Unsupervised Person Re-identification

“…Our CAP [11] constructs a proxylevel memory bank to perform intra-and inter-camera contrastive learning at proxy-level. Later on, ICE [12] and Liu et al [35] boost SPCL and CAP via augmenting the models with instance-level contrastive learning, while MGH [15] and Isobe et al [31] integrate contrastive learning with other techniques such as hypergraph or Fourier augmentation. In contrast to them [12], [15], [17], [31], [35], this work extends CAP via sticking on the proxy-level contrastive learning alone to keep the approach simple yet effective.…”

“…3) Effectiveness of the strategies in online association: In the design of our online association, we propose an instanceproxy balanced similarity and a camera-aware nearest neighbor criterion to select positive proxies. In order to validate their effectiveness, we compare them, respectively, with the original instance-to-proxy similarity and the global KNN that are ordinarily used [6], [15]. Table III presents the comparison results.…”

“…For instance, RLCC [16] refines pseudo labels via clustering consensus over consecutive training generations. MGH [15] adopts a hypergraph to propagate and refine pseudo labels generated by global DBSCAN, intra-camera DBSCAN, and global KNN. HCT [20] utilizes hierarchical clustering and Zheng et al [18] propose an online clustering strategy to generate high quality pseudo labels.…”

“…The comparison results are summarized in Table IV and Table V. 1) Comparison with purely unsupervised methods on person Re-ID: 11 representative or recent purely unsupervised methods are included for comparison. Most of these methods are clustering-based, and CAP [11] (our preliminary work), ICE [12], MGH [15], MGCE-HCL [17], and Liu et al [35] also adopt the contrastive learning technique same as this work. Both ICE and MGH are built upon CAP.…”

“…Existing methods commonly resort to pseudo labels for learning. Clustering [5], [20], k-NN [21], [22], graph [23], [24], or hypergraph [15] based techniques have been developed to generate pseudo labels. Clustering-based methods such as BUC [5] and HCT [20] conduct learning in a camera-agnostic way, which can capture the similarity within IDs but neglect the intra-ID variance mostly caused by the change of camera views.…”