Diffusion Transformers (DiTs) have demonstrated remarkable generative capabilities, particularly benefiting from Transformer architectures that enhance visual and artistic fidelity. However, their inherently sequential denoising process results in high inference latency, limiting their deployment in real-time scenarios. Existing training-free acceleration approaches typically reuse intermediate features at fixed timesteps or layers, overlooking the evolving semantic focus across denoising stages and Transformer blocks.To address this, we propose Sortblock, a training-free inference acceleration framework that dynamically caches block-wise features based on their similarity across adjacent timesteps. By ranking the evolution of residuals, Sortblock adaptively determines a recomputation ratio, selectively skipping redundant computations while preserving generation quality. Furthermore, we incorporate a lightweight linear prediction mechanism to reduce accumulated errors in skipped blocks.Extensive experiments across various tasks and DiT architectures demonstrate that Sortblock achieves over 2 times inference speedup with minimal degradation in output quality, offering an effective and generalizable solution for accelerating diffusion-based generative models.

WiFi indoor localization plays an important role in many real-world applications and has gained widespread attentions from both academia and industry over the past decade. While existing works have already achieved remarkable performance under various practical scenarios, they solely utilize multiple WiFi Access Points (APs) to achieve a higher accuracy and do not deeply explore the intrinsic relationships among them. To further unleash the potential of multiple APs and reach the limits of WiFi indoor localization, in this paper, we propose Qidi, a novel multi-APs collaboration based localization system. To the best of our knowledge, we are the first to explicitly reveal three underlying relationships among multiple spatially distributed APs, i.e., consistency, continuity, and non-uniformity. By combining these three basic principles with the unique characteristics of specific tasks, a series of long-lasting practical challenges, such as automatic phase offset calibration, bilateral angle ambiguity, elevation angle estimation with Uniform Linear Array (ULA) and Non-Line-of-Sight (NLoS), can be efficiently resolved. Extensive experiments in various complex environments are provided to demonstrate that Qidi can achieve 2.4°, 3.2° median errors of joint azimuth and elevation angle estimation, and 0.4mlocalization median error even in 20m×20mexhibition hall. Moreover, a one-month longitudinal evaluation conducted on a real-world deployed WiFi ISAC system further validates the effectiveness and robustness of the proposed localization system.

WiFi-based indoor localization, a fundamental technology for numerous real-world applications, including indoor navigation and emergency evacuation, has garnered significant attention over the past decade. While existing works have already achieved remarkable performance under various practical scenarios, they simply take corridor as a test scenario, and attribute performance degradation to complex multipath. To further release the functionality of corridor as bridges connecting different physical spaces, in this paper, we propose Widor, the firstWiFi indoor localization system that designed specifically for corridor. We first explore the neglected elevation angle dimension information in WiFi positioning and creatively place AP vertically to avoid the large-angle effect, which is caused by the unique slender structure of the corridor. To eliminate the impact of complex multipath introduced by the narrow environment, we make full use of the Toeplitz structure of the covariance matrix to further improve the spatial resolution of existing commercial WiFi AP without increasing the additional hardware cost. Furthermore, we design a tailored multi-APs joint height compensation algorithm to bridge the gap between 2-D and 3-D localization, which iteratively optimize the height difference between the AP and the client through an alternating optimization method. Both the time dimensional information and map constraint can be used to further improve localization accuracy. We evaluate Widor under various complex corridors in an$82m \times 65m$building, and extensive experimental results show that Widor can achieve$5.2^{\circ}$median angle estimation error and 58 cm localization median error. We also highlight that the proposed Widor system has little impact on the communication performance of existing commercial APs, and can be easily extended to any slender building scenarios, such as mines and tunnels, further broadening the application boundaries of WiFi indoor localization.

WiFi-based indoor localization serves as a critical foundation for numerous real-world applications and has attracted widespread attentions over the past decade. Recent advancements have demonstrated the feasibility of achieving decimeter-level accuracy by leveraging Angle of Arrival (AoA) information. However, existing commercial WiFi Access Points (APs) suffer from phase offset across different antennas, which significantly degrade the performance of AoA-based methods. Previous works either relied on labor-intensive manual calibration or involved inaccurate and non-robust automatic calibration, which hinders their widespread use in large-scale deployments. Moreover, to expand the signal coverage and enhance communication performance, the inter-antenna spacing in existing commercial APs typically exceeds the standard half-wavelength. The resulting angle ambiguity problem can mislead target detection results, which has not been well resolved in existing works. To address the above two practical challenges, in this paper, we propose CAFE, a practical WiFi indoor localization system based on theContinuousAngleFocusingEffect. The key insight lies on the fact that the angle information of multiple APs originates from the same client, and thus exhibits highly convergent properties in both the temporal and spatial dimensions. By further exploring the binary nature of phase offset and the periodicity of grating lobes, our approach can efficiently resolve the above two practical challenges. Extensive experiments are provided to demonstrate the effectiveness of the proposed CAFE system, which outperforms state-of-the-art methods by$22.1\%$in median localization error for simple scenarios and by$37.1\%$for complex multipath scenarios.

In the above article [1], on page 13900, right column, there is an empty reference citation “[?]” in the sentence “By applying Model-Agnostic Meta-Learning (MAML) to fingerprint localization, MetaLoc [?] enables the model to quickly adapt to new environments based on the obtained meta-parameters, thus reducing human labor costs.” The missing reference is listed below as [2].

The YOLO series has become the most popular frame-work for real-time object detection due to its reasonable trade-off between speed and accuracy. However, we observe that the speed and accuracy of YOLOs are negatively affected by the NMS. Recently, end-to-end Transformer-based detectors (DETRs) have provided an alternative to eliminating NMS. Nevertheless, the high computational cost limits their practicality and hinders them from fully exploiting the advantage of excluding NMS. In this paper, we propose the Real-Time DEtection TRansformer (RT-DETR), the first real-time end-to-end object detector to our best knowledge that addresses the above dilemma. We build RT-DETR in two steps, drawing on the advanced DETR: first we focus on maintaining accuracy while improving speed, followed by maintaining speed while improving accuracy. Specifically, we design an efficient hybrid encoder to expeditiously process multi-scale features by decoupling intra-scale interaction and cross-scale fusion to improve speed. Then, we propose the uncertainty-minimal query selection to provide high-quality initial queries to the decoder, thereby improving accuracy. In addition, RT-DETR supports flexible speed tuning by adjusting the number of decoder layers to adapt to various scenarios without retraining. Our RT-DETR-R50 /R101 achieves 53.1% 154.3% AP on COCO and 108 /74 FPS on T4 GPU, outperforming previously advanced YOLOs in both speed and accuracy. Furthermore, RT-DETR-R50 outperforms DINO-R50 by 2.2% AP in accuracy and about 21 times in FPS. After pre-training with Objects365, RT-DETR-R50 / R101 achieves 55.3% 156.2% AP. The project page: https://zhao-yian.github.io/RTDEtr.

The mutual interference mitigation between identical or similar radar systems in autonomous driving has gained wide spread attention from both academia and industry. The resulted ghost target interference will reduce the sensitivity of the radar sensor and increase the false alarm rate. To tackle this problem, in this paper, we make full use of two characteristics of interference to achieve ghost target interference mitigation in the Doppler domain. The key insight lies in the fact that the interference is one-way propagation, and thus the resulted ghost target can be converted to the noise floor in the Doppler domain through random slow-time coding. Moreover, the high power characteristic of interference allows us to further enhance the interference mitigation performance by adopting a signal-to-interference-plus-noise ratio (SINR) maximization principle. Numerical examples are provided to demonstrate the effectiveness of the proposed interference mitigation approach.

Recent advancements in WiFi indoor localization have demonstrated the potential for achieving decimeter-level accuracy based on Angle of Arrival (AoA). However, existing commercial WiFi Access Points (APs) suffer from phase offset across different antennas, which significantly degrade the performance of AoA-based methods in practical deployment. Previous work either relied on labor-intensive manual calibration or involved inaccurate and non-robust automatic calibration. In this paper, we propose AutoCali, an accurate and robust automatic phase offset calibration system. The key insight is to utilize the binary nature of phase offsets and the property that triangulation exhibits higher convergence when the correct combination of phase offsets is employed. Extensive experiments demonstrate that AutoCali outperforms state-of-the-art methods by 22.1% in median localization error for simple scenarios and by 37.1% for complex multipath scenarios.

Air quality control is essential for assessing the impact on human health, environment and climate. Single-particle mass spectrometry (SPMS) is a powerful measurement tool for providing the chemical composition of air-transported particle matter (PM) in real-time. Common methods to classify PM according to characteristic ion patterns in their mass spectra are based on clustering methods which generally require manual postprocessing and are not suitable for real-time automated air quality monitoring. A number of automated classification models trained on labeled SPMS data were proposed recently by the authors. As it appeared, the most advanced methods of them, based on deep-learning convolutional neural networks (CNN), still have difficulties in distinguishing between classes of similar but distinctive mass spectra. In this work, we propose a novel fuzzy convolutional neural network (FCNN) combining fuzzy network and CNN to accurately classify particle mass spectral patterns. FCNN models integrate the respective advantages of fuzzy and neural networks to effectively separate non-isolated and overlapping features of similar patterns through fuzzy information, and also inherently optimize the fuzzy rule parameters through NN back propagation. To validate the performance of FCNN, a benchmark dataset with 37,406 samples in 13 particle classes was created. Compared to CNN, with FCNN 10 out of 13 classes could be classified with higher accuracy, especially those distinguishing subtle differences in the mass spectra. Applied to automated SPMS analysis, the proposed FCNN tackles the classification challenges posed by the chemical complexity of aerosol particles and opens up ways to foster the development of specific, real-time air quality monitoring and pollution source identification systems.

WiFi-based indoor localization has gained widespread attention due to the pervasive availability of WiFi Access Points (APs). While signal processing-based methods can achieve decimeter-level localization, their performance is constrained by the limited spatial resolution of WiFi systems, especially in complex environments with strong interference. By contrast, deep learning-based methods have achieved impressive performance even in complex environments, which however often fail to generalize to new environments. In this paper, we propose a novel framework to learn domain-invariant model for WiFi-based indoor localization, which maintains impressive performance across different environments. The key insight is to design a deep learning-based WiFi localization system through the perspective of signal processing. Specifically, we let the neural network estimate APs-centered polar coordinates to avoid fitting the coordinates of APs strongly correlated with the environment, enabling us to obtain the domain-invariant model. To unleash the potential of neural networks in regressing high-precision parameters, we design a beamforming layer to integrate the knowledge of signal processing. Furthermore, we propose a multi-task learning scheme to further improve localization accuracy. Extensive experiments on diverse datasets have demonstrated that the localization performance of our method outperforms state-of-the-art methods and demonstrates superiority under cross-domain conditions.

Garbage collection in SSDs causes write amplification. The key to mitigating this problem is separating data by lifetime. Prior works proposed using machine learning to accurately predict data lifetime but prediction is performed at the host side, burdening the host storage stack. We present PHFTL, a practical, holistic FTL design with device-side learning-based data separation. The machine learning model in PHFTL accurately and adaptively predicts the lifetime of every written page. A suite of enabling techniques are introduced to keep computation and storage overhead low. Extensive evaluation of PHFTL demonstrates superiority over state-of-the-art and feasibility on real hardware.