Deep Metric Learning with Angular Loss

By Jian Wang, Feng Zhou, Shilei Wen, Xiao Liu, Yuanqing Lin. 2017

548 citations recorded by the source corpus.

The modern image search system requires semantic understanding of image, and a key yet under-addressed problem is to learn a good metric for measuring the similarity between images. While deep metric learning has yielded impressive performance gains by extracting high level abstractions from image data, a proper objective loss function becomes the central issue to boost the performance. In this paper, we propose a novel angular loss, which takes angle relationship into account, for learning better similarity metric. Whereas previous metric learning methods focus on optimizing the similarity (contrastive loss) or relative similarity (triplet loss) of image pairs, our proposed method aims at constraining the angle at the negative point of triplet triangles. Several favorable properties are observed when compared with conventional methods. First, scale invariance is introduced, improving the robustness of objective against feature variance. Second, a third-order geometric constraint is inherently imposed, capturing additional local structure of triplet triangles than contrastive loss or triplet loss. Third, better convergence has been demonstrated by experiments on three publicly available datasets.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

Semantic Drift Compensation for Class-Incremental Learning (2020, methodology): “They are typically used for image retrieval [40], face recognition [33],etc.”
Relational Knowledge Distillation (2019, background): “Note that 2 normalization for embedding is widely used in deep metric learning [7, 13, 21, 22, 34, 41].”
Learning an Evolutionary Embedding via Massive Knowledge Distillation (2020, background): “Recently, the angular constraints (Liu et al. 2016, 2017; Wang et al. 2017) have been introduced into deep embedding learning.”
Multi-Similarity Loss With General Pair Weighting for Deep Metric Learning (2019, background): “…includes contrastive loss [6], triplet loss [10], triplet-center loss [8], quadruplet loss [18], lifted structure loss [25], N-pairs loss [29], binomial deviance loss [40], histogram loss [32], angular loss [34], distance weighted margin-based loss [38], hierarchical triplet loss (HTL) [4], etc .”
Doodle to Search: Practical Zero-Shot Sketch-Based Image Retrieval (2019, methodology): “pes [26]) etc. This domain shift issue is further addressed by crossdomain deep learning-based methods [27, 37], where they have used classical ranking losses, such as, contrastive loss, triplet loss [32] or more elegant HOLEF loss [30] within a siamese like network. Based on the problem at hand, tw”
Representation Learning for Visual-Relational Knowledge Graphs (2017, background): “There is a large body of work on metric learning where the objective is to generate image embeddings such that a pairwise distance-based loss is minimized [38,39,40,41,42].”
Metric Learning for Adversarial Robustness (2019, methodology): “In this paper, we define the embedding distance between two examples using the angular distance [42]: D(h(x (i) a ), h(x (j) p,n)) = 1− |h(x a )·h(x (j) p,n))| ||h(x (i) a )||2||h(x (j) p,n))||2 , where we choose to encode the information in the angular metric space.”
Revisiting Training Strategies and Generalization Performance in Deep Metric Learning (2020, unknown): “Angular (Wang et al., 2017) 62.”
Boosting Adversarial Training with Hypersphere Embedding (2020, background): “However, due to the high computational burden induced by metric learning, different variants of HE are presented and become popular on face recognition tasks (Wang et al., 2017a; Liu et al., 2017).”
Multi-Attention Multi-Class Constraint for Fine-grained Image Recognition (2018, background): “nvergence. To alleviate this issue, the N-pair loss [42] is introduced to consider multiple negative samples in training, and exhibits higher eciency and performance. More recently, the angular loss [43] enhances N-pair loss by integrating high-order constraint that captures additional local structu”
Classification is a Strong Baseline for Deep Metric Learning (2018, background): “Recent approaches in metric learning design the loss function to consider the relationships of all the samples within the training batch[9, 17, 18, 21, 27], and achieve state-of-the-art performance”
Deep Learning Based Vulnerability Detection: Are We There Yet? (2020, background): “learning, to create amaximal separation between classes [43], [44].”
RepMet: Representative-based metric learning for classification and one-shot object detection (2018, background): “distance between the samples in an embedding space generated by a neural network. The basic loss function for training such an embedding is the triplet loss [41], or one of its recent generalizations [34,35,39]. These losses are designed to make the embedding space semantically meaningful, such that”
ProxyNCA++: Revisiting and Revitalizing Proxy Neighborhood Component Analysis (2020, background): “In addition to contrastive and triplet loss, there is a long line of work which proposes new loss functions, such as angular loss [31], histogram loss [26], margin-based loss [33], and hierarchical triplet loss [6].”
Deep Metric Learning: A Survey (2019, methodology): “Part loss value was calculated for each part of an image divided into five parts to evaluate each of the local patterns as a separate part Loss functions are one of the most important parts of the deep metric learning models.”
Triplet-Center Loss for Multi-view 3D Object Retrieval (2018, background): “Several variants of triplet loss have also been proposed, such as [19, 23, 37].”
Metric Learning for Novelty and Anomaly Detection (2018, unknown): “Several other losses have been proposed for Siamese networks [10, 23, 26, 29, 30] but in this paper we will evaluate results with the contrastive loss to provide a simple baseline on which to improve.”
Characterizing Generalization under Out-Of-Distribution Shifts in Deep Metric Learning (2021, background): “These ranking tasks can also leverage additional context such as geometrical embedding structures [69, 8].”