Comparator Networks

By Weidi Xie, Li Shen, Andrew Zisserman. 2018

77 citations recorded by the source corpus.

The objective of this work is set-based verification, e.g. to decide if two sets of images of a face are of the same person or not. The traditional approach to this problem is to learn to generate a feature vector per image, aggregate them into one vector to represent the set, and then compute the cosine similarity between sets. Instead, we design a neural network architecture that can directly learn set-wise verification. Our contributions are: (i) We propose a Deep Comparator Network (DCN) that can ingest a pair of sets (each may contain a variable number of images) as inputs, and compute a similarity between the pair--this involves attending to multiple discriminative local regions (landmarks), and comparing local descriptors between pairs of faces; (ii) To encourage high-quality representations for each set, internal competition is introduced for recalibration based on the landmark score; (iii) Inspired by image retrieval, a novel hard sample mining regime is proposed to control the sampling process, such that the DCN is complementary to the standard image classification models. Evaluations on the IARPA Janus face recognition benchmarks show that the comparator networks outperform the previous state-of-the-art results by a large margin.

Subfields: Computer Vision and Pattern Recognition, Model Architectures, Signal Processing

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

In Model Architectures

In Signal Processing

Citations in the corpus

CrossTransformers: spatially-aware few-shot transfer (2020, background): “Part-based correspondences for recognition [103] have been particularly successful in fine-grained face retrieval and recognition [48, 95].”
ArcFace: Additive Angular Margin Loss for Deep Face Recognition (2018, methodology): “tches. On the IJB-B and IJB-C datasets, we employ the VGG2 dataset as the training data and the ResNet50 as the embedding network to train ArcFace for the fair comparison with the most recent methods [6,42,41]. In Table7, we compare the TAR (@FAR=1e-4) of ArcFace with the previous stateof-the-art mo”
Data Uncertainty Learning in Face Recognition (2020, unknown): “lly use non-reﬁned MegaFace. Method Training Data IJB-C (TPR@FPR) 0:001% 0:01% 0:1% AUC Yin et al. [48] 0.5M - - 69:3 - Cao et al. [5] 3.3M 74:7 84:0 91:0 - Multicolumn [46] 3.3M 77:1 86:2 92:7 - DCN [45] 3.3M - 88:5 94:7 - PFE [35] 4.4M 89:64 93:25 95:49 - Baseline 3.6M 83:06 92:16 95:83 91:97 PFEr”
Discriminability Distillation in Group Representation Learning (2020, unknown): “-C verication benchmark. The True Accept Rates (TAR) vs. False Accept Rate (FAR) are reported Method IJB-C (TAR@FAR) 1e-6(%) 1e-5(%) 1e-4(%) 1e-3(%) 1e-2(%) Yin et al. [63] - - - 0.1 83.8 Xie et al. [59] - - 88.5 94.7 98.3 Zhao et al. [66] - 82.6 89.5 93.5 96.2 multicolumn [60] - 77.1 86.2 92.7 96.8”
Probabilistic Face Embeddings (2019, unknown): “protocol) and IJB-A. Method Training Data IJB-C (TAR@FAR) 0:001% 0:01% 0:1% 1% Yin et al. [49] 0.5M - - 69:3 83:8 Cao et al. [3] 3.3M 74:7 84:0 91:0 96:0 Multicolumn [45] 3.3M 77:1 86:2 92:7 96:8 DCN [44] 3.3M - 88:5 94:7 98:3 Baseline 4.4M 70:10 85:37 93:61 96:91 PFE fuse 4.4M 83:14 92:38 95:47 97:”
CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition (2020, methodology): “5 exhibits the performance of different methods, e.g. , Multicolumn [35], DCN [34], Adacos [38], P2SGrad [39], PFE [25] and MV-Arc-Softmax [31] on IJB-B and IJB-C 1 : 1 veriﬁcation, our method again achieves the best performance.”
ShrinkTeaNet: Million-scale Lightweight Face Recognition via Shrinking Teacher-Student Networks (2019, unknown): “ods on 1:1 IJB-B and IJB-C Veriﬁcation protocol. The accuracy is reported at TAR (@FAR=1e-4). Method Training Data IJB-B IJB-C SENet50 [51] VGG2 0.800 0.840 MN-VC [52] VGG2 0.831 0.862 ResNet50 + DCN [53] VGG2 0.841 0.880 ArcFace [42] VGG2 0.898 0.921 ArcFace [42] MS1M 0.942 0.956 MV1 [32] MS1M 0.90”
Voxceleb: Large-scale speaker verification in the wild (2020, background): “It is inspired by the relation networks and their use in face comparisons (Santoro et al., 2017; Xie et al., 2018).”
Sub-center ArcFace: Boosting Face Recognition by Large-Scale Noisy Web Faces (2020, methodology): “To facilitate comparisons, we abbreviate different settings by the experiment number (E*) in the table and only focus on the TAR@FAR=1e-4 of IJB-C, which is more ob-jective and less affected by the noise within the test data [38].”