Proxy Anchor Loss for Deep Metric Learning

Abstract

Existing metric learning losses can be categorized into two classes: pair-based and proxy-based losses. The former class can leverage fine-grained semantic relations between data points, but slows convergence in general due to its high training complexity. In contrast, the latter class enables fast and reliable convergence, but cannot consider the rich datato-data relations. This paper presents a new proxy-based loss that takes advantages of both pair-and proxy-based methods and overcomes their limitations. Thanks to the use of proxies, our loss boosts the speed of convergence and is robust against noisy labels and outliers. At the same time, it allows embedding vectors of data to interact with each other through its gradients to exploit data-to-data relations. Our method is evaluated on four public benchmarks, where a standard network trained with our loss achieves state-ofthe-art performance and most quickly converges.

Adaptive Deep Metric Ensemble Learning with Consensus

“…We compared many state-of-the-art alternatives, including four ensemble methods, i.e., ABE (Attention-Based Ensem-ble) [28], DREML (Deep Randomized Ensemble for Metric Learning) [27], DCES (Divide-and-Conquer Embedding Space) [29], and A-BIER (Adversarial-Boosting Independent Embeddings Robustly) [11], as well as typical methods, e.g., Triplet [13], HTL (Hierarchical Triplet Loss) [15] and Margin [20]. Besides, proxy-based approaches like Proxy-NCA [18] and Proxy-Anchor [19] while MS (Multi-Similarity loss) [17] and its variant DR-MS (Direction Regularized MS) [33], EE-MS (Embedding Expansion MS) [21] are compared. In addition, the records of ALA (Adaptive Learnable Assessment) [22] and HDML (Hardness-aware Deep Metric Learning) [23] are reported.…”

“…Hence, we obtain the consensus-aware adaptive ensemble, which are employed for yielding the embedding representations of test samples across a wide range of classes. Following [19,36,37], we use Recall@h(R@h) and mAP@R for image retrieval task and Normalized Mutual Information (NMI) [29] for clustering task. R@h is the ratio of queries which have at least one sample within the class of query in the leading h items retrieved by model, and mAP@R is a more strict criterion which considers the sequence of items.…”

“…The majority of them belong to pair-based functions, e.g., contrastive loss [6,12], triplet loss [7,13], N-pair loss [14], hierarchical triplet loss [15], ranked list loss [16], and multi-similarity loss with general pair weighting [17]. Besides, some loss functions adopt proxy mechanism, such as proxy Neighborhood Component Analysis (ProxyNCA) [18] and proxy anchor [19], to speedup the convergence of model training, where the optimization is carried out on the proxies of triplets, each of which involves an anchor point and similar/dissimilar proxy points. In addition, histogram loss [3] was explored for learning deep embedding by estimating two distributions of similarities for positive and negative sample pairs.…”

Deep Semi-Supervised Metric Learning Via Identification of Manifold Memberships

“…We then generate and maintain a set of L 2 -normalized proxies {p c } C c=0 , p c ∈ R d , where d is the dimension of the embeddings z, and we want data with confident pseudolabels to be be drawn close to proxies of the same class. We feed the pseudolabels and confidence values we obtained to a loss function adapted from [14]. Let…”

“…In this paper, we compare our results against state-of-theart semi-supervised algorithms SDEC [9] and PULLMH [5]; fully-supervised algorithms Proxy-Anchor [14], SoftTriple [19] and ProxyNCA++ [20]; and unsupervised algorithm DEC [21].…”