mmWave radar-based human pose estimation is garnering increasing attention in Internet of Things (IoT) applications owing to its robustness under adverse lighting conditions and occlusion scenarios. However, accurately estimating human poses from mmWave signals remains challenging due to two key limitations: the lack of structural consistency when perceiving targets as a unified whole, and the absence of spatiotemporal consistency when tracking individual targets. To tackle these two challenges, this paper introduces RF-PoseR, a dedicated human pose rectifier for mmWave radar-based pose estimation. RF-PoseR serves as a corrective framework that refines the pose outputs generated by mmWave pose estimation systems. Based on the observation that perceptual inconsistencies in radar signals result in severely erroneous joint estimations, we introduce spatial and temporal constraints to enhance pose estimation accuracy. Specifically, RF-PoseR incorporates three core components: (1) unreliable joint removal guided by spatial structure, (2) pose completion leveraging temporal continuity, and (3) cross-modal pre-training utilizing large-scale visual pose datasets. Extensive experiments on mmWave radar pose datasets demonstrate that RF-PoseR significantly enhances the accuracy of poses generated by existing radar-based estimation networks. Furthermore, experiments on visual datasets confirm the method’s broader applicability beyond radar perception tasks.

Millimeter-wave (mmWave) radar systems with large array have pushed radar sensing into a new era, thanks to their high angular resolution. However, our long-term experiments indicate that array elements exhibit phase drift over time and require periodic phase calibration to maintain high-resolution, creating an obstacle for practical high-resolution mmWave sensing with large array. Unfortunately, existing calibration methods are inadequate for periodic recalibration, either because they rely on artificial references or fail to provide sufficient precision. To address this challenge, we introduce AutoCalib, the first framework designed to automatically and accurately calibrate high-resolution mmWave radars by identifying Ambient Radio Anchors (ARAs)—naturally existing objects in ambient environments that offer stable phase references. AutoCalib achieves calibration by first generating spatial spectrum templates based on theoretical electromagnetic characteristics. It then employs a pattern-matching and scoring mechanism to accurately detect these anchors and select the optimal one for calibration. Extensive experiments across 11 environments demonstrate that AutoCalib is capable of identifying ARAs that existing methods miss due to their focus on strong reflectors. AutoCalib's calibration performance approaches corner reflectors (74% phase error reduction) while outperforming existing methods by 83%. Beyond radar calibration, AutoCalib effectively supports other phase-dependent applications like handheld imaging, delivering 96% of corner reflector calibration performance without artificial references.

Human tracking plays a crucial role in various wireless sensing applications. However, recent advancements have primarily focused on constrained experimental scenarios with less interference, often involving a few individuals performing actions in an empty space without obstacles. In empirical unconstrained scenarios, such as daily office scenes, severe interference and attenuation caused by chaotic environments is inevitable which results in dramatic performance degradation. In this paper, we introduce TBDNet, which incorporates tracking-before-detect (TBD) from conventional signal processing into learning-based models, achieving impressive tracking performance in unconstrained scenarios. TBDNet follows first-track-then-detect pipeline. It maps input heatmap sequence into high-level frame-wise features to adapt the time-varying intensity distribution and motion pattern of targets. After that, the temporal information is accumulated in feature space to obtain trace proposals. We then predict the accurate positions and probability of traces at each timestamp. To assess the efficiency of TBDNet, we collect and release the first RF-UNIT (RF-based Unconstrained Indoor Tracking) dataset, which comprises 4,030,880 radar heatmaps and the corresponding tracking annotations under 6 different scenarios. To our knowledge, RF-UNIT is the first dataset for RF-based human tracking in unconstrained scenes. We anticipate that TBDNet and the RF-UNIT dataset will significantly contribute to the advancement of RF-based sensing technologies.

Recent developments in X-ray image based pulmonary nodule detection have achieved remarkable results. However, existing methods are focused on transferring off-the-shelf coarse-grained classification models and fine-grained detection models rather than developing a dedicated framework optimized for nodule detection. In this paper, we propose PN-DetX, which as we know is the first dedicated pulmonary nodule detection framework. PN-DetX incorporates feature fusion and self-attention into X-ray based pulmonary nodule detection tasks, achieving improved detection performance. Specifically, PN-DetX adopts CSPDarknet backbone to extract features, and utilizes feature augmentation module to fuse features from different levels followed by context aggregation module to aggregate semantic information. To evaluate the efficacy of our method, we collect a LArge-scale Pulmonary NOdule Detection dataset, LAPNOD, comprising 2954 X-ray images along with expert-annotated ground truths. Experiments demonstrates that our method outperforms baseline by 3.8 mAP and 5.1 AP50. The dataset and codes will be made in public.

Pulmonary nodule detection in chest X-rays facilitates early lung cancer diagnosis, yet is constrained by the lack of labeled data requiring expertise and resources for annotation. In this study, we propose ReCo-CXR, the first pre-training framework specifically designed for pulmonary nodule detection. This framework incorporates image-level and region-level contrastive learning, leveraging 112,120 unlabeled chest X-ray images to enhance model representations and initializations. Specifically, image-level contrastive learning is employed to extract global translation-invariant features. Furthermore, we propose region-level contrastive learning to capture position-sensitive features by maximizing the similarity of corresponding regions across different views. Experimental results on the clinically-aligned LAPNOD and public NODE21 datasets demonstrate that pre-training with ReCo-CXR consistently improves the performance of various detection models, achieving best AP0.5improvements of 2.9% and 9.0%, respectively. Our code and pre-trained weights will be made publicly available.

Toxic gases inhalation is the most common cause of death in fire scenes, which can make people unconscious and unable to save themselves. Hence, discovering the unconscious victims is crucial to improve their survival rate. In this paper, we propose RF-Search, a victim searching system with RF device mounted on the drone. The challenge mainly comes from the fact that drone motion would overwhelm the subtle vital signs utilized for victim identification. To resolve this problem, we have noted that the physical signature of drone motion has been encoded in stationary object reflections. Leveraging this unique physical signature, we propose to identify the unconscious victim through the spatio-temporal correlation between signals reflected from the victim and the surrounding stationary objects. To extract respiration information of the victim, we propose a motion segmentation module and a motion compensation module to suppress the signal variation caused by drone movement. Extensive experiments have demonstrated that our system could achieve an accuracy of 92.5% for victim identification.

A novel structure-aware unsupervised network is proposed to deal with low-light image enhancement issues based on the inspiration of Retinex theory and self-supervised perceptual loss. It comprises four main components, namely the original structural similarity module, the novel color consistency module, the attentional enhancement module, and the naturalness discriminator module. Specially for the structural similarity module, an embedded structural feature extractor (SFE) model capable of generating structure correspondence is well designed and pre-trained by employing the contrastive learning technique, and a multi-scale structural similarity distance is introduced to optimize the SFE network. Besides, a self-supervised color consistency module is established by using a degraded estimation algorithm for recovering the missing colors. The whole enhancement framework operates in unsupervised manners and finally obtains the best naturalness image quality evaluator metric. Experimental results demonstrate that the proposed unsupervised network is able to recover natural structure and color images more effectively, which would also help to enlarge the practical application without collecting paired datasets in advance.

The Exemplar Memory (EM) design has shown its effectiveness in facilitating the unsupervised person re-identification (RE-ID). However, there are obvious defects in the update strategies with most existing results, such as the inability to eliminate static errors and ensure convergence stability of learning. To address these issues, in this paper, we propose a novel center feature learning scheme to improve the update strategies of the traditional EM design for unsupervised RE-ID problems. First, the EM module is regarded as a center feature of a cluster of images, then the goal is transformed into pulling the similar images close to while pushing the dissimilar images away from the center feature space. Second, in order to provide effective guidelines on reducing static errors, we propose an error-memory module to improve the central feature learning performances. In addition, an error-prediction module is designed as well to ensure the stability of convergence. Besides, a camera-invariance learning strategy is also introduced to further improve the proposed algorithm. Finally, extensive comparative experiments are conducted on Market-1501 and DukeMTMC-reID datasets to demonstrate the effectiveness and improvements of the proposed method over existing results. The code of this work is available at https://github.com/binquanwang/CFL_master.

Person re-identification (re-ID) aims to identify the same persons' images across different cameras by deploying a trained re-ID model. Nevertheless, the domain bias between different datasets can lead to considerably low identification accuracy when directly applying a trained re-ID model from one dataset to another. Therefore, in order to develop a kind of more transferable re-ID methods with strong robustness, this paper proposes a fast momentum contrast learning framework, which consists of a teacher encoder and a fast momentum encoder, from a reinforcing visual representation learning point of view. The fast momentum encoder produced by the teacher encoder is employed to build a dynamic dictionary and then design the contrast learning loss function together with the teacher encoder. In addition, both the individual loss and the clustering loss are adopted for obtaining the optimal performances in unsupervised person re-ID models. The proposed framework achieves one stage of end-to-end training and experimental results demonstrate that our identification accuracy outperforms existing advanced methods.