Online multi-object tracking (MOT) often fails under frequent occlusions and non-linear motion because corrupted temporal appearance embeddings introduce association ambiguity and identity switches (IDS). Existing methods mainly improve backend association, but largely overlook the direct restoration of degraded features themselves. We propose SGFR-Track, an uncertainty-gated spectral feature restoration framework for online MOT. The key idea is to treat occlusion-induced feature corruption as high-frequency temporal noise and restore discriminative appearance cues before association. To this end, we first introduce a Gating Funnel based on the Bidirectional Matching Margin (BMM) to diagnose tracklet ambiguity online and activate restoration only for high-risk tracklets, improving efficiency while avoiding over-smoothing of reliable samples. We then design a Spectral-Guided Feature Restoration (SGFR) module that performs adaptive denoising in the frequency domain via the Discrete Cosine Transform (DCT). Extensive experiments under the unified BoxMOT framework on DanceTrack, MOT17, and MOT20 show that SGFR-Track consistently improves overall tracking performance while preserving real-time speed. These results demonstrate that spectral signal processing offers an effective and physically interpretable direction for enhancing feature robustness in online MOT.

Recent text-to-image diffusion models have shown significant advancements in high-quality image synthesis. Represented by Stable Diffusion, these models have been applied in various fields, such as art design, digital development, personalization, etc. However, they still fail to generate images that properly comply with the given text prompts, especially when encountering multiple subjects. Current approaches for multi-subject synthesis focus on latent code optimization at specific timesteps,yet yielding compromised text-image alignment. In this paper, we explore the generation of diffusion models through the lens of both noise initialization and denoising process, and propose an effective two-stage Attention Refinement and Inheritance (ARI) approach. Firstly, we update the initial noise by refining the cross-attention maps between different subject tokens, employing strategies of separation, energization, and centralization. Then, we introduce attention-level interpolation as an inheritancemechanism to explicitly regulate the denoising process. This enables the image generation to maintain high correspondence with the refined initial attention, thereby achieving a more controllable synthesis. Extensive experiments show that our approach achieves more faithful alignment between generated images and text prompts compared with state-of-the-art multi-subject image synthesis methods.

Electrocardiogram (ECG) remains the most common tool for detecting cardiovascular disease (CVD). The emergence of deep learning has significantly improved the efficiency and accuracy of automated ECG-based diagnosis. However, existing deep learning models often fail to integrate the three pillars of a cardiologist's diagnosis: temporal dynamics, spatial relationships, and patient-specific context. To address these limitations, we propose Hetero-Cardio, a novel framework designed to emulate this holistic diagnostic reasoning. Hetero-Cardio employs a parallel dual-pathway architecture, and a Transformer pathway captures long-range temporal dependencies across the cardiac cycle. Meanwhile, a heterogeneous graph neural network (HGNN) is employed to integrate spatial and patient-specific features. The core innovation lies in this heterogeneous design, which enables the model to learn patient-conditioned spatial features that reflect how inter-lead relationships should be interpreted differently based on individual context. By synergistically fusing temporal, spatial, and contextual information, Hetero-Cardio learns a comprehensive diagnostic representation. Our model achieves highly competitive performance on classification tasks using two large-scale public datasets, PTB-XL and CPSC 2018. Our code is publicly available at: https://github.com/Cryostars/HeteroCardio

With the development of diffusion-based customization methods like DreamBooth, individuals now have access to train the models that can generate their personalized images. Despite the convenience, malicious users have misused these techniques to create fake images, thereby triggering a privacy security crisis. In light of this, proactive adversarial attacks are proposed to protect users against customization. The adversarial examples are trained to distort the customization model's outputs and thus block the misuse. In this paper, we propose DisDiff (Disrupting Diffusion), a novel adversarial attack method to disrupt the diffusion model outputs. We first delve into the intrinsic image-text relationships, well-known as cross-attention, and empirically find that the subject-identifier token plays an important role in guiding image generation. Thus, we propose the Cross-Attention Erasure module to explicitly "erase" the indicated attention maps and disrupt the text guidance. Besides, we analyze the influence of the sampling process of the diffusion model on Projected Gradient Descent (PGD) attack and introduce a novel Merit Sampling Scheduler to adaptively modulate the perturbation updating amplitude in a step-aware manner. Our DisDiff outperforms the state-of-the-art methods by 12.75% of FDFR scores and 7.25% of ISM scores across two facial benchmarks and two commonly used prompts on average.