In Defense of the Triplet Loss for Person Re-Identification

By Alexander Hermans, Lucas Beyer, Bastian Leibe. 2017

3,600 citations recorded by the source corpus.

In the past few years, the field of computer vision has gone through a revolution fueled mainly by the advent of large datasets and the adoption of deep convolutional neural networks for end-to-end learning. The person re-identification subfield is no exception to this. Unfortunately, a prevailing belief in the community seems to be that the triplet loss is inferior to using surrogate losses (classification, verification) followed by a separate metric learning step. We show that, for models trained from scratch as well as pretrained ones, using a variant of the triplet loss to perform end-to-end deep metric learning outperforms most other published methods by a large margin.

Subfields: Computer Vision and Pattern Recognition, 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 Signal Processing

Citations in the corpus

Parameter-Efficient Transfer Learning for NLP (2019, methodology): “This technique attains state-of-the-art performance on many vision tasks, including classification (Kornblith et al., 2018), fine-grained classifcation (Hermans et al., 2017), segmentation (Long et al., 2015), and detection (Girshick et al., 2014).”
S4L: Self-Supervised Semi-Supervised Learning (2019, methodology): “Following [16], we implement Lu as the batch hard triplet loss [13] with a soft margin.”
Unsupervised Embedding Learning via Invariant and Spreading Instance Feature (2019, background): “Triplet (hard) is the online hard negative sample within each batch for training [16], and the margin parameter is set to 0.”
Revisiting Self-Supervised Visual Representation Learning (2019, methodology): “[8] proposes to use the triplet loss [40, 18] in order to scale this pretext task to a large number of images (hence, classes) present in the ImageNet dataset.”
Random Erasing Data Augmentation (2017, methodology): “For TriNet and SVDNet, we use the same model as propose in [8] and [21], and follow the same training strategy to them.”
VoxCeleb2: Deep Speaker Recognition (2018, methodology): “Pair-wise losses such as the contrastive loss are notoriously difficult to train [41], and hence to avoid suboptimal local minima early on in training, we proceed in two stages: first, pre-training for identifcation using a softmax loss, then, second, fine-tuning with the contrastive”
CLIMATE-FEVER: A Dataset for Verification of Real-World Climate Claims (2020, methodology): “We apply hard positive and negative mining [18, 7] to compensate for the large number of possible negative examples.”
Contrastive Adaptation Network for Unsupervised Domain Adaptation (2019, neutral): “face recognition [6], person re-identification [15], etc.”
Continual Representation Learning for Biometric Identification (2020, background): “Biometric identification [23,51], including face recognition [11, 37, 67] and person re-identification (re-id) [34, 66, 71], has achieved significant progress in the recent years due to the advances in modern learnable representations [8, 9, 11, 20, 34, 52, 60, 67] and emerging large datasets [17, 2”
Fine-Grained Action Retrieval Through Multiple Parts-of-Speech Embeddings (2019, background): “Recently, neural networks trained with a ranking loss considering image pairs [27], triplets [35], quadruplets [5] or beyond [32], have been considered for metric learning [17, 35] and for a broad range of search tasks such as face/person identiﬁcation [29, 5, 16, 2] or instance retrieval [10, 27].”
Classifying, Segmenting, and Tracking Object Instances in Video with Mask Propagation (2019, neutral): “Model Classification cls head Mask R-CNN [16], ResNeXt-101 32x48d [30] cls head Localization bbox head Mask R-CNN [16] bbox head Segmentation mask head DeepLabv3 [9], Box2Seg [29] mask head Tracking tracking head UnOVOST [47], ReID Net [18, 31] mask propagation head Optical Flow PWC-Net [35] -”
Deep learning in video multi-object tracking: A survey (2019, methodology): “[113] directly used the 128-d feature vector extracted by the ReID triplet CNN proposed in [114], and combined it with other appearance-based features (as an alternative to an appearance-less version of the algorithm).”
RGB-Infrared Cross-Modality Person Re-Identification via Joint Pixel and Feature Alignment (2019, methodology): “Existing methods can be grouped into hand-crafted descriptors [22, 36, 19, 34], metric learning methods [43, 16, 20] and deep learning algorithms [42, 11, 28, 21, 44, 30, 15, 35, 33, 18, 17, 1, 31].”
Joint Discriminative and Generative Learning for Person Re-Identification (2019, methodology): “Current state-of-the-art re-id methods widely formulate the tasks as deep metric learning problems [13, 55], or use classiﬁcation losses as the proxy targets to learn deep embeddings [23, 39, 41, 49,54,57].”
Generalizing a Person Retrieval Model Hetero- and Homogeneously (2018, methodology): “For similarity learning, we employ the triplet loss used in [42], which is formulated as, where X represents images in a training batch, x a is an anchor point. x p is a hardest (farthest) sample in the same class with x a , and x n is a hardest (closest) sample of a diﬀerent class to x a . m is a…”
Exploit the Unknown Gradually: One-Shot Video-Based Person Re-identification by Stepwise Learning (2018, methodology): “In recent years, person re-ID on the large-scale video data, such as surveillance videos, has attracted significant attention [10, 20, 28, 32, 35].”
FairMOT: On the Fairness of Detection and Re-identification in Multiple Object Tracking (2020, background): “, 2017) propose to crop the image regions of the detections and feed them to re-ID networks (Zheng et al., 2017b; Hermans et al., 2017; Luo et al., 2019a) to extract image features.”
Thin-Plate Spline Motion Model for Image Animation (2022, methodology): “We use the same pre-trained re-identification networks for bodies [12] and faces [1] as MRAA [28] to extract identities from the generated and driving images.”
Deep Metric Learning: A Survey (2019, background): “Deep metric learning for person re-identification provides us to use end-to-end learning between the input image and the transformed feature space [40].”