Noise-robust re-identification with triple-consistency perception

State-of-the-art person re-identification (ReID) models use Convolutional Neural Networks (CNN) for feature extraction and comparison. Often these models fail to recognize all the intra- and inter-class variations that emerge in person ReID, making it harder to discriminate between data subjects. In this paper we seek to reduce these problems and improve performance by combining two state-of-the-art models. We use the Omni-Scale Network (OSNet) as our CNN to test the Market1501 and DukeMTMC-ReID datasets for person ReID. To fully utilize the potential of these datasets, we apply the spatialtemporal constraint which extracts the camera ID and timestamp from each image to form a distribution. We combine these two methods to create a hybrid model titled Spatial-Temporal OmniScale Network (st-OSNet). Our model attains a Rank-1 (R1) accuracy of 98.2% and mean average precision (mAP) of 92.7% for the Market1501 dataset. For the DukeMTMC-reID dataset our model achieves 94.3% R1 and 86.1% mAP, hereby surpassing the results of OSNet by a large margin for both datasets (94.3%, 86.4%, 88.4%, 76.1%, respectively).

Person re-identification (Re-ID) aims to retrieve a person of interest across multiple nonoverlapping cameras. In the last few years, the construction of aerial person Re-ID datasets has been appealing because visual surveillance through unmanned aerial vehicle (UAV) platforms has become very valuable in real-world scenarios. However, great differences are exhibited between the pedestrian images captured by ground cameras and those captured by UAVs. Person Re-ID methods based on ground person images have difficulty when performing Re-ID based on aerial person images. In this paper, we first propose a novel meta-transfer learning method for person Re-ID in aerial imagery; this approach trains a generalisable Re-ID model to learn discriminative feature representations for aerial person images. Specifically, a meta-learning strategy is introduced to study a feature extractor, and a transfer learning strategy is introduced to utilise and further improve the acquired meta-knowledge. To overcome the convergence speed and recognition accuracy reductions caused by the presence of difficult categories in the given dataset, we propose a learning strategy based on curriculum sampling that is harmonised with our meta-transfer learning framework. In addition, a new metric formulation of sample similarity is introduced based on Mahalanobis distance to improve the optimisation of the model. Extensive comparative evaluation experiments are conducted on the large-scale aerial Re-ID dataset, and the results obtained show that our method achieves a Rank-1 accuracy of 63.63 $$\%$$ and a mean average precision (mAP) of 38.02 $$\%$$ , demonstrating its potential for completing person Re-ID in aerial images. Ablation studies also validate that each component contributes to improving the performance of the model.

This paper proposes a weakly supervised cross-domain person re-identification (Re-ID) method based on small sample data. In order to reduce the cost of data collection and annotation, the model design focuses on extracting and abstracting the information contained in the data under limited conditions. In this paper, we focus on the problems of strong data dependence, weak cross-domain capability and low accuracy in Re-ID in weakly supervised scenarios. Our contributions are as follows: first, we implement a joint training framework with the help of small sample learning and cross-domain migration for Re-ID. Second, with the help of residual compensation and fusion attention module, the RCFA module is designed, and the model framework is built on this basis to improve the cross-domain ability of the model. Third, to solve the problem of low accuracy caused by insufficient data coverage of small samples, a fusion of shallow features and deep features is designed to enable the model to weighted fusion of shallow detail information and deep semantic information. Finally, by selecting different camera images in Market1501 dataset and DukeMTMC-reID dataset as small samples, respectively, and introducing another dataset data for joint training, we demonstrate the feasibility of this joint training framework, which can perform weakly supervised cross-domain Re-ID based on small sample data.

Hierarchical fusion and local-aware transformer for occluded person re-identification.

Although great progress in supervised person re-identification (Re-ID) has been made recently, due to the viewpoint variation of a person, Re-ID remains a massive visual challenge. Most existing viewpoint-based person Re-ID methods project images from each viewpoint into separated and unrelated sub-feature spaces. They only model the identity-level distribution inside an individual viewpoint but ignore the underlying relationship between different viewpoints. To address this problem, we propose a novel approach, called Viewpoint-Aware Loss with Angular Regularization (VA-reID). Instead of one subspace for each viewpoint, our method projects the feature from different viewpoints into a unified hypersphere and effectively models the feature distribution on both the identity-level and the viewpoint-level. In addition, rather than modeling different viewpoints as hard labels used for conventional viewpoint classification, we introduce viewpoint-aware adaptive label smoothing regularization (VALSR) that assigns the adaptive soft label to feature representation. VALSR can effectively solve the ambiguity of the viewpoint cluster label assignment. Extensive experiments on the Market1501 and DukeMTMC-reID datasets demonstrated that our method outperforms the state-of-the-art supervised Re-ID methods.

Person re-identification is widely regarded as an image retrieval problem. Given a pedestrian-of-interest (query) image in one camera, the person re-identification system aims to identify the pictures of the same person from an image pool (gallery). Due to the differences in camera pixels, pose, illumination, occlusion, and intra-class variations across different cameras, the task of person re-identification remained challenging to the community of computer vision scientists. In this paper, we propose a multi-task network, based on a uniform partition network, which computes the identification loss and verification loss of two input images simultaneously. Given a pair of images as input, the system predicts the identities of the two input images and outputs a similarity score at the same time, to indicate whether they belong to the same identity or not. To get more fine-grained part-level features, we adopted the part-based convolutional baseline network for feature extraction of each input image and output a convolutional descriptor consisting of six local features. Our model achieved 81.19% mAP and 93.34% rank-1 accuracy on Market-1501 datasets. It also achieved 72.12% mAP and 85.59% rank-1 accuracy on DukeMTMC-reID. Comparing them with those of state-of-the-art, our model outperformed the state-of-the-art by a margin of 3.79% mAP, 1.03% rank-1, and 6.02% mAP, 3.79% rank-1 on Market-1501 and DukeMTMC-reID, respectively.

Person re-identification (Re-ID) using infrared surveillance cameras has attracted increasing attention due to its robustness under low-light conditions. However, infrared images generally suffer from a low spatial resolution, which degrades Re-ID performance. To address this issue, this study proposes a part attention and contrastive loss-based super-resolution reconstruction network (PCSR-Net) and a unified infrared-only Re-ID framework. The proposed PCSR-Net consists of a correlation-based super-resolution reconstruction network (CoSR-Net), a feature extractor for Re-ID, and a part attention mechanism that estimates the importance of different body regions. In addition, contrastive loss and part-aware reconstruction loss are incorporated to guide the super-resolution process toward identity-discriminative representations. Experimental results on DBPerson-Recog-DB1 and SYSU-MM01 demonstrate that the proposed method outperforms state-of-the-art approaches in terms of the equal error rate (EER), mean average precision (mAP), and rank-1 accuracy, validating its effectiveness for infrared-based person Re-ID.

A review on video person re-identification based on deep learning

Animal re-identification (Re-ID) has recently gained substantial attention in the AI research community due to its high impact on biodiversity monitoring and unique research challenges arising from environmental factors. The subtle distinguishing patterns like stripes or spots, handling new species and the inherent open-set nature make the problem even harder. To address these complexities, foundation models trained on labeled, large-scale and multi-species animal Re-ID datasets have recently been introduced to enable zero-shot Re-ID. However, our benchmarking reveals significant gaps in their zero-shot Re-ID performance for both known and unknown species. While this highlights the need for collecting labeled data in new domains, exhaustive annotation for Re-ID is laborious and requires domain expertise. Our analyses also show that existing unsupervised (USL) and active learning (AL) Re-ID methods underperform for animal Re-ID. To address these limitations, we introduce a novel AL Re-ID framework that leverages complementary clustering methods to uncover and target structurally ambiguous regions in the embedding space for mining pairs of samples that are both informative and broadly representative. Oracle feedback on these pairs, in the form of must-link and cannot-link constraints, facilitates a simple annotation interface, which naturally integrates with existing USL methods through our proposed constrained clustering refinement algorithm. Through extensive experiments, we demonstrate that, by utilizing only 0.033% of all possible annotations, our approach consistently outperforms existing foundational, USL and AL baselines. Specifically, we report an average improvement of 10.49%, 11.19% and 3.99% (mAP) on 13 wildlife datasets over foundational, USL and AL methods, respectively, while attaining state-of-the-art performance on each dataset. Furthermore, we also show an improvement of 11.09%, 8.2% and 2.06% (AUC ROC) for unknown individuals in an open-world setting. We also present results on 2 publicly available person Re-ID datasets, showing average gains of 7.96% and 2.86% (mAP) over existing USL and AL Re-ID methods.

Two-phase self-supervised pretraining for object re-identification

Traditional person re-identification methods are mostly based on supervised state of manually annotated data sets. Although the recognition accuracy is high, the algorithm completely relies on effective annotation information and lacks practicability. Based on the above problems, an unsupervised person re-identification method is proposed, which refines features to guide multi-label distribution sorting learning. Firstly, a single and multi-label classification network structure is constructed to improve the matching accuracy of tags by learning the characteristics of each person and its k nearest neighbor and sharing the same class. Secondly, the refinement feature extraction module is designed and embedded into the original ResNet50 network, and the potential key areas in the person image are adaptively located. While weakening the appearance differences of the features, multi-information mining is carried out to obtain the refined features of the characters and assist the classification network to improve the performance. Thirdly, reliable multi-category labels are screened by punishing the sorting between positive and negative categories of labels in multi-category labels. Finally, the network is supervised by combining multi-label distribution sorting loss and multi-label classification loss. Experiments show that without any annotation information, the recognition accuracy of Market-1501 and DukeMTMC-reID datasets is 58.4% and 42.3%, respectively, and the mAP of Market-1501 dataset is increased by 12.9%.

Person reidentification (Re-ID) aims to match specific pedestrians across nonoverlapping camera views. Due to the dramatic disparities between different datasets, transferring a Re-ID model trained on the source domain to the target domain is challenging. Some outstanding unsupervised domain adaptation (UDA) Re-ID methods use clustering to generate pseudolabels, optimizing the model on the target domain, but the pseudolabels are inevitably noisy. To address the above issues, we propose a framework named mini-transformer with pooling (MTP) to facilitate the generation of superior quality pseudolabels by improving the model’s feature representation capability. First, we introduce an effective mini-transformer (MT) that can be placed directly behind the CNNs as a feature extractor to capture long-range dependency. Then, we design two delicate pooling methods named global hybrid pooling (GHP) and global subvalue pooling (GSVP) to suit mini-transformer’s tremendous capability without increasing computational complexity. Specifically, GHP can keep more global information and GSVP can keep more discriminative information. Finally, experiments on four mainstream UDA Re-ID tasks demonstrate that MTP achieves competitive mAP and rank-1 accuracy to the current state-of-the-art methods, suggesting that our technique is simple but effective. In addition to UDA Re-ID, our MTP can be extended to other supervised retrieval tasks.

Most zero-shot learning (ZSL) methods based on generative adversarial networks (GANs) utilize random noise and semantic descriptions to synthesize image features of unseen classes, which alleviates the problem of training data imbalance between seen and unseen classes. However, these methods usually only learn the distributions of seen classes in the training stage, ignoring the unseen ones. Due to the different distributions of seen and unseen samples, i.e., image features, these methods cannot generate unseen features of sufficient quality, so the performances are also limited, especially for the generalized zero-shot learning (GZSL) setting. In this article, we propose a general transductive method based on GANs, called GT-GAN, which can improve the quality of generated unseen image features and therefore benefit the classification. A new loss function is introduced to make the relative positions between each unseen image and its <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula> nearest neighbors in the feature space as consistent as possible with their relative positions in the semantic space; this loss function may be easily applied in most existing GAN-based models. Experimental results on five benchmark datasets show a significant improvement in accuracy compared with that of original models, especially in the GZSL setting.

Person re-identification (Re-ID) is crucial in intelligent surveillance, requiring precise identification of individuals across multiple camera viewpoints. Traditional distance-based methods, such as Euclidean and Cosine, struggle with challenges like posture variations and occlusions, limiting their effectiveness. This study explores deep metric learning models, specifically Siamese and Triplet networks, to improve Re-ID performance. We evaluate these methods on the Market-1501 dataset using Cumulative Matching Characteristic (CMC) and Cumulative Distribution Function (CDF) curves. Our findings reveal that the Triplet network outperforms traditional approaches at higher ranks, achieving Rank-5 accuracy of 78.6% and Rank-10 accuracy of 93%, while its Rank-1 accuracy remains low (0.06%). In contrast, Euclidean and Cosine distances show poor Rank-1 performance (2% and 0.30%, respectively), highlighting their limitations. Additionally, incorporating VGG16 enhances feature extraction, improving recognition by capturing fine-grained spatial details. This comparative study highlights the effectiveness of deep metric learning and underscores its potential for real-world surveillance applications. However, the computational demands of deep networks present challenges for real-time deployment. Future research should focus on optimizing model efficiency, reducing computational costs, and extending evaluations to real-time scenarios.

The performance of deep neural networks highly depends on the quality and volume of the training data. However, cost-effective labelling processes such as crowdsourcing and web crawling often lead to data with noisy (i.e., wrong) labels. Making models robust to this label noise is thus of prime importance. A common approach is using loss distributions to model the label noise. However, the robustness of these methods highly depends on the accuracy of the division of training set into clean and noisy samples. In this work, we dive in this research direction highlighting the existing problem of treating this distribution globally and propose a class-conditional approach to split the clean and noisy samples. We apply our approach to the popular DivideMix algorithm and show how the local treatment fares better with respect to the global treatment of loss distribution. We validate our hypothesis on two popular benchmark datasets and show substantial improvements over the baseline experiments. We further analyze the effectiveness of the proposal using two different metrics - Noise Division Accuracy and Classiness.