Quadruplet Loss Research Articles

Deep Quadruplet Network for Hyperspectral Image Classification with a Small Number of Samples

This study proposes a deep quadruplet network (DQN) for hyperspectral image classification given the limitation of having a small number of samples. A quadruplet network is designed, which makes use of a new quadruplet loss function in order to learn a feature space where the distances between samples from the same class are shortened, while those from a different class are enlarged. A deep 3-D convolutional neural network (CNN) with characteristics of both dense convolution and dilated convolution is then employed and embedded in the quadruplet network to extract spatial-spectral features. Finally, the nearest neighbor (NN) classifier is used to accomplish the classification in the learned feature space. The results show that the proposed network can learn a feature space and is able to undertake hyperspectral image classification using only a limited number of samples. The main highlights of the study include: (1) The proposed approach was found to have high overall accuracy and can be classified as state-of-the-art; (2) Results of the ablation study suggest that all the modules of the proposed approach are effective in improving accuracy and that the proposed quadruplet loss contributes the most; (3) Time-analysis shows the proposed methodology has a similar level of time consumption as compared with existing methods.

Pedestrian Retrieval Using Generated Samples and Multistream Layer in Sensor Networks

In this paper, we propose a novel loss function named the hybrid quadruplet loss (HQL) to utilize the generated samples for pedestrian retrieval in sensor networks. The proposed HQL employs a set of quadruplets in order to maintain an appropriate margin between the real sample and the generated sample, reduce the intra-class variations and enlarge the inter-class variations. By this way, the generalization of the deep model could be improved. Furthermore, to identify the extremely similar pedestrians, we propose a novel multistream layer to mine imperceptible information from different aspects. The proposed multistream layer utilizes various filters with different morphologies to capture discriminative features in multiple scales, and it is flexible to follow any convolutional layer. Experiments on the three large-scale pedestrian retrieval databases (Market1501, CUHK03, and DukeMTMC-reID) demonstrate that the proposed method outperforms other state-of-the-art methods.

A Deep Quadruplet Network for Local Descriptor Learning

Generating a distinguishable feature for local patch is a main task in computer vision which aims at matching local patches. Recently, local patch descriptors from deep convolutional neural network (CNN) with a triplet loss have achieved promising performance. In this paper, we design a quadruplet loss, which can achieve a better result than other pairwise loss and triplet loss methods. Our loss is inspired by the thoughts of uniform distribution. It separates non-matching examples by using the hard sampled non-matching pairs in a batch, and simultaneously uses the random sampled non-matching examples to keep non-matching pairs to obey uniform distribution. A compact descriptor named QuadrupletNet is generated by combining the proposed quadruplet loss and L2Net CNN architecture. From our experiment, QuadrupletNet shows better performance on the Brown dataset and Hpatches dataset than Triplet loss methods on the same training set. The pre-trained QuadrupletNet is publicly available.

ClothingNet: Cross-Domain Clothing Retrieval With Feature Fusion and Quadruplet Loss

Cross-domain clothing retrieval is an active research topic because of its massive potential applications in fashion industry. Due to the large number of garment categories or styles, and different clothing appearances caused by different camera angles, different shooting conditions, different messy background environments, or different postures of the dressed human body, the retrieval accuracy of traditional consumer-to-shop scheme is always low. In this paper, based on the framework of deep convolution neural network, a novel cross-domain clothing retrieval method is proposed by using feature fusion and quadruplet loss function, which is named as ClothingNet. First, the pre-trained deep neural network Resnet-50 is adopted to extract feature map of clothing images. The extracted high-level image features can thus be merged with middle-level features, and the final feature representation of clothing images can be obtained by constraining the fusion feature values to a certain range in term of L2 norm. This fusion feature provides a comprehensive description of the differences between clothing images. For effectively training our ClothingNet, the cross-domain clothing images are organized in form of a quadruplet for calculating its loss function, and the network parameters can be optimized according to back propagation scheme via stochastic gradient descent of loss function. Our proposed method is validated on two public datasets for clothing retrieval, DARN and DeepFashion, showing that the top-50 retrieval accuracy is 35.67% and 53.52% respectively. Experimental results illustrate the effectiveness of our clothing retrieval method.

Learning Domain-Invariant Discriminative Features for Heterogeneous Face Recognition

Heterogeneous face recognition (HFR), referring to matching face images across different domains, is a challenging problem due to the vast cross-domain discrepancy and insufficient pairwise cross-domain training data. This article proposes a quadruplet framework for learning domain-invariant discriminative features (DIDF) for HFR, which integrates domain-level and class-level alignment in one unified network. The domain-level alignment reduces the cross-domain distribution discrepancy. The class-level alignment based on a special quadruplet loss is developed to further diminish the intra-class variations and enlarge the inter-class separability among instances, thus handling the misalignment and adversarial equilibrium problems confronted by the domain-level alignment. With a bidirectional cross-domain data selection strategy, the quadruplet loss-based method prominently enriches the training set and further eliminates the cross-modality shift. Benefiting from the joint supervision and mutual reinforcement of these two components, the domain invariance and class discrimination of identity features are guaranteed. Extensive experiments on the challenging CASIA NIR-VIS 2.0 database, the Oulu-CASIA NIR&VIS database, the BUAA-VisNir database, and the IIIT-D viewed sketch database demonstrate the effectiveness and preferable generalization capability of the proposed method.

Joint Learning of Generative Translator and Classifier for Visually Similar Classes

In this paper, we propose a Generative Translation Classification Network (GTCN) for improving visual classification accuracy in settings where classes are visually similar and data is scarce. For this purpose, we propose joint learning from a scratch to train a classifier and a generative stochastic translation network end-to-end. The translation network is used to perform on-line data augmentation across classes, whereas previous works have mostly involved domain adaptation. To help the model further benefit from this data-augmentation, we introduce an adaptive fade-in loss and a quadruplet loss. We perform experiments on multiple datasets to demonstrate the proposed method's performance in varied settings. Of particular interest, training on 40% of the dataset is enough for our model to surpass the performance of baselines trained on the full dataset. When our architecture is trained on the full dataset, we achieve comparable performance with state-of-the-art methods despite using a light-weight architecture.

Deep Quadruplet Appearance Learning for Vehicle Re-Identification

Vehicle re-identification (Re-ID) plays an important role in intelligent transportation systems. It usually suffers from various challenges encountered on the real-life environments, such as viewpoint variations, illumination changes, object occlusions, and other complicated scenarios. To effectively improve the vehicle Re-ID performance, a new method, called the deep quadruplet appearance learning (DQAL), is proposed in this paper. The novelty of the proposed DQAL lies on the consideration of the special difficulty in vehicle Re-ID that the vehicles with the same model and color but different identities (IDs) are highly similar to each other. For that, the proposed DQAL designs the concept of quadruplet and forms the quadruplets as the input, where each quadruplet is composed of the anchor (or target), positive, negative, and the specially considered high-similar (i.e., the same model and color but different IDs with respect to the anchor) vehicle samples. Then, the quadruplet network with the incorporation of the proposed quadruplet loss and softmax loss is developed to learn a more discriminative feature for vehicle Re-ID, especially discerning those difficult high-similar cases. Extensive experiments conducted on two commonly used datasets VeRi-776 and VehicleID have demonstrated that the proposed DQAL approach outperforms multiple recently reported vehicle Re-ID methods.

Quadruplet-based deep hashing for image retrieval

Hashing methods have been widely used for large-scale image retrieval. Learning deep hashing networks with a pairwise loss or a triplet ranking loss has become a common framework. The pairwise loss and triplet ranking loss, respectively, focus on preserving the pairwise similarity and the relative similarity ordering. In this paper, we design a quadruplet loss that can fully explore the similarity relation between image pairs to decrease the intraclass variation and increase the interclass variation. Moreover, we propose a deep architecture based on quadruplet loss and optimal adaptive margins for learning hash codes. Extensive experimental results show that our method achieves state-of-the-art performance on several benchmark image retrieval datasets.

Zero-Shot Cross-Media Embedding Learning With Dual Adversarial Distribution Network

Existing cross-media retrieval methods are mainly based on the condition where the training set covers all the categories in the testing set, which lack extensibility to retrieve data of new categories. Thus, zero-shot cross-media retrieval has been a promising direction in practical application, aiming to retrieve data of new categories (unseen categories), only with data of limited known categories (seen categories) for training. It is challenging for not only the heterogeneous distributions across different media types, but also the inconsistent semantics across seen and unseen categories need to be handled. To address the above issues, we propose dual adversarial distribution network (DADN) , to learn common embeddings and explore the knowledge from word-embeddings of different categories. The main contributions are as follows. First, zero-shot cross-media dual generative adversarial networks architecture is proposed, in which two kinds of generative adversarial networks (GANs) for common embedding generation and representation reconstruction form dual processes. The dual GANs mutually promote to model semantic and underlying structure information, which generalizes across different categories on heterogeneous distributions and boosts correlation learning. Second, distribution matching with maximum mean discrepancy criterion is proposed to combine with dual GANs, which enhances distribution matching between common embeddings and category word-embeddings. Finally, adversarial inter-media metric constraint is proposed with an inter-media loss and a quadruplet loss, which further model the inter-media correlation information and improve semantic ranking ability. The experiments on four widely used cross-media datasets demonstrate the effectiveness of our DADN approach.

Non-Rigid 3D Model Retrieval Based on Quadruplet Convolutional Neural Networks

Non-rigid 3-D model retrieval is a challenging problem in 3-D shape analysis. Recently, deep learning-based 3-D feature extraction methods have been studied and have achieved better performance than the previous state-of-the-art methods. Inspired by the quadruplet neural networks proposed for learning local image feature descriptors, we propose a novel non-rigid 3-D model retrieval method based on quadruplet convolutional neural networks. For training the proposed networks, the quadruplet samples are first selected using the online sampling method. For each 3-D model, the wave kernel signature descriptor of each vertex is computed, and its corresponding multi-energy shape distribution matrix is constructed as the input of the network. Then, the quadruplet convolutional neural networks are trained using our improved quadruplet loss function, which not only preserves the advantages of existing quadruplet loss functions but also decreases the risk of underfitting. For the query sample, the 3-D shape features are computed using one branch of the trained quadruplet networks. Finally, the retrieval results are obtained by the L2 distance measure. Extensive experimental results have validated the effectiveness of the proposed method.