Multi-Label Continual Learning (MLCL) is a framework designed for class-incremental multi-label image recognition. However, MLCL faces two critical challenges: the construction of label relationships on past-missing and future-missing partial labels of training data, and the problem of catastrophic forgetting , which leads to poor generalization. To address these challenges, this study proposes an enhanced version of the Augmented Graph Convolutional Network (AGCN++), capable of constructing cross-task label relationships and mitigating catastrophic forgetting. First, an Augmented Correlation Matrix (ACM) is constructed across all observed classes, incorporating intra-task relationships derived from hard label statistics. Additionally, inter-task relationships are established by leveraging both hard and soft labels obtained from the data, as well as a constructed expert network. Next, a novel partial label encoder (PLE) is introduced for MLCL, enabling the extraction of dynamic class representations for each partial label image as graph nodes. This PLE also facilitates the generation of soft labels, which contribute to the creation of a more persuasive ACM and effectively mitigate forgetting. Lastly, a relationship-preserving constrainter is proposed to address the issue of forgetting label dependencies across old tasks. In the AGCN++, the label relationships topology can be augmented automatically, thereby generating efficient class representations. The effectiveness of the proposed method is evaluated using two multi-label image benchmarks. The experimental results demonstrate that the proposed approach is highly effective in the context of MLCL image recognition. It can establish compelling correlations across tasks, even in scenarios where the old task labels are missing.

Vehicle-to-everything (V2X) perception is an innovative technology that enhances vehicle perception accuracy, thereby elevating the security and reliability of autonomous systems. However, existing V2X perception methods exhibit limitations at traffic intersections due to a failure to account for effective collaboration between dynamic vehicles and stable Roadside Unit (RSU) in adapting to traffic environments. To harness the use of RSU and adapt V2X perception models to dynamic scenes, we propose an Adaptive Road-to-Vehicle Perception (AR2VP) approach, building V2X perception based on road-to-vehicle visual information. In AR2VP, we leverage RSU to offer stable, wide-range sensing capabilities and serve as communication hubs. AR2VP is devised to tackle both intra-scene and inter-scene changes. For the intra-scene changes, we construct a Dynamic Perception Representation (DPR) module, which effectively integrates vehicle perceptions, enabling vehicles to capture a more comprehensive range of dynamic factors within the scene. Moreover, we introduce a Road-to-Vehicle Perception Compensating (R2VPC) module, aiming at preserving the maximized RSU perception information in the presence of intra-scene changes. For the inter-scene changes, we implement a Prompt-Replay mechanism leveraging the RSU's limited storage capacity to retain a subset of typical historical scene samples, maintaining model robustness in response to inter-scene changes. We conduct comprehensive experiments on 3D object detection and segmentation, and the results show that AR2VP excels in both performance-bandwidth trade-offs and adaptability within dynamic environments. Our code is available at https://github.com/tjy1423317192/AR2VP .

Earphones have become a popular voice input and interaction device. However, airborne speech is susceptible to ambient noise, making it necessary to improve the quality and intelligibility of speech on earphones in noisy conditions. As the dual-microphone structure (i.e., outer and in-ear microphones) has been widely adopted in earphones (especially ANC earphones), we design EarSpeech which exploits in-ear acoustic sensory as the complementary modality to enable airborne speech enhancement. The key idea of EarSpeech is that in-ear speech is less sensitive to ambient noise and exhibits a correlation with airborne speech. However, due to the occlusion effect, in-ear speech has limited bandwidth, making it challenging to directly correlate with full-band airborne speech. Therefore, we exploit the occlusion effect to carry out theoretical modeling and quantitative analysis of this cross-channel correlation and study how to leverage such cross-channel correlation for speech enhancement. Specifically, we design a series of methodologies including data augmentation, deep learning-based fusion, and noise mixture scheme, to improve the generalization, effectiveness, and robustness of EarSpeech, respectively. Lastly, we conduct real-world experiments to evaluate the performance of our system. Specifically, EarSpeech achieves an average improvement ratio of 27.23% and 13.92% in terms of PESQ and STOI, respectively, and significantly improves SI-SDR by 8.91 dB. Benefiting from data augmentation, EarSpeech can achieve comparable performance with a small-scale dataset that is 40 times less than the original dataset. In addition, we validate the generalization of different users, speech content, and language types, respectively, as well as robustness in the real world via comprehensive experiments. The audio demo of EarSpeech is available on https://github.com/EarSpeech/earspeech.github.io/.

Recently, plenty of millimeter-wave image concealed object detection models have achieved superior performance on benchmark datasets. The success of these models heavily depends on a large amount of fully-annotated data. However, noisy-label data often exist in a large-scale dataset, which deteriorate the performance of models. In this paper, we delve into the training of the concealed object detector with label noise, where we recognize divergent types of noisy annotations. To tackle the multi-type label noise problem, we propose a unified framework including two stages: the label noise modeling and the correction training. We design the region-level label noise modeling, which analyzes and models label noise via the classification branch loss curves. To correct the labels during training, we design the uncertainty-weighted loss for label correction. The experiments on a millimeter-wave security image dataset demonstrate the effectiveness of our proposed training framework, which can alleviate the negative influence of noisy annotations on network training. By introducing our framework, the object detection network can achieve a prominent performance improvement on mean average precision (0.176 on average) compared to the one trained with noisy labels.

Point-cloud-based place recognition is a key component for outdoor large-scale Simultaneous Localization And Mapping (SLAM) in re-localization. However, most methods have limited generalization ability for unseen environments. To address this issue, a Hybrid Voxel- and Point- wise network, named HVP-Net, is proposed. This network utilizes sparse convolutions to learn the local detail of the voxel- wise features and proposed lightweight grouped efficient attention mechanisms to capture the global representations of the point- wise features. To enhance the discrimination of the global descriptors, these two kinds of features are fused in an interactive way to take advantage of point- wise features without information loss and voxel- wise ones robust to local noises. In addition, a positive-ranking guided triplet loss is proposed, which further considers the consistency of distance ranking between different anchor-positive pairs in both Euclidean and feature space. Experiments on the benchmark, KITTI, NCLT, and one self-collected dataset show that HVP-Net achieves state-of-the-art results and can effectively improve the generalization ability for unseen environments.