Fast adversarial training is susceptible to a failure mode known as catastrophic overfitting (CO), which can cause a sudden drop in the model’s robust accuracy to zero against multi-step attacks. Previous research has proposed various approaches to mitigate this problem, typically comparing the model’s state before and after CO or contrasting it with the multi-step adversarial training process to identify the inducing factors. However, these approaches analyze only a limited set of factors, and the robustness of the proposed solutions has a significant gap to that of multi-step adversarial training. In this paper, we propose a novel framework called Neighborhood Diversity-Aligned Fast Adversarial Training (NDA-FAT), designed to systematically address the issue throughout the training process. NDA-FAT commences by employing a novel neighborhood diversity random initialization strategy to generate multiple starting points within the vicinity of samples, thereby enhancing the diversity of subsequent adversarial examples. We then propose an adversarial effectiveness weighting method, which is used to select high-quality starting points to generate effective adversarial examples for training. Moreover, we propose a min-max alignment regularization approach aimed at improving the model’s local linearity, thereby further enhancing the model’s robustness. Evaluation results on multiple benchmark datasets demonstrate that NDA-FAT effectively mitigates the catastrophic overfitting problem and significantly improves the model’s robustness.

Person reidentification matches person observations captured by nonoverlapping cameras at different times and locations. Due to the appearance variations caused by view angle, pose, lighting, background, and occlusion, person reidentification is a challenging task. To improve the accuracy of person reidentification, this paper proposes a metric learning method. The main idea of our method is to learn a low-dimensional metric space, in which the features extracted from different observations of the same person are pulled, and the features extracted from observations of different persons but in the neighborhood of original feature space are pushed. Then all the features are projected to this metric space for matching. Compared with other metric learning methods, our method differentiates negative samples according to their importances in constructing the metric space, and pays more attention to restrain the more confusable negative pairs. Moreover, our method can simultaneously learn a uniform metric space for multiple features that is practical for multifeature fusion, an effective way to improve reidentification performance. Experimental results on three popular data sets show that our method is more robust against appearance changes and it outperforms other state-of-the-art reidentification methods.

In surveillance videos, the pictures of a same person often present significant variation which makes person re-identification difficult. Though the globe appearances may present great difference, some local patches still have great similarities, and human eyes can be used to distinguish the identity of each person via these local patches. Inspired from it, patch matching is introduced in person re-identification and has been shown to be an efficient method to solve these problems caused by different viewpoints, poses, camera settings, illumination and occlusion. But until now there is no guide for how to decide the size of patch, and most researches got these patches either by dense sampling or coarse partition. In both case, the person structure information is missing. To improve re-identification accuracy, we propose a method via person DPM (Deformable Parts Model) partition. First, both compared appearances are partitioned into several body parts by pre-trained person DPM and these parts are grouped according to their positions in the body; Second, part matching is conducted between two appearances' parts in each group based on deep learning features; Finally, fusing the similarities of each group are to decide whether these two appearances are from the same person or not. Experiments on VIPeR dataset illustrate that without supervised training, the proposed method can obtain good re-identification performances compared with state-of-the-art methods.