Visual tracking plays an essential role in autonomous unmanned aerial vehicle (UAV) systems. However, existing trackers still face a challenging trade-off between tracking accuracy and inference speed, which hinders their deployment on resource-constrained onboard platforms. To address this issue, we present a novel automated network architecture search framework to explore robust and efficient visual object trackers. Specifically, we first design a multi-scale cosine cross-attention unit to enhance feature interaction, and a computation-efficient self-attention unit to improve feature extraction. Based on the above two units and convolution operators, we then design a hierarchical search space to provide multi-dimensional search. Finally, we design a hardware-aware search strategy with heatmap distillation to automatically search for the optimal tracker within the specific search space. Heatmap distillation is applied to enhance the performance of potential architectures, while the hardware-aware mechanism guides the architecture search to satisfy UAV platform constraints. Extensive experiments on multiple UAV tracking benchmarks demonstrate that the searched tracker achieves an average success score of 64.6 with a low inference latency of 32.5 ms per frame. Compared to the representative one-stream method MixViT, our searched tracker achieves approximately 1.8× higher inference speed. These results highlight the potential of automated hardware-aware model design for developing high-performance visual trackers under practical deployment constraints.
Traditional source separation methods rely on coarse-grained categorical labels through labeling all vocals collectively without distinguishing individual voices in an audio mixture, which inherently limits the ability to isolate single tracks. While fine-grained annotations could partially address this issue, they demand substantial resources and face challenges in extracting tracks from raw signals. To overcome these limitations, we propose to extract each track through decomposing the patterns of data generation. Specifically, we refine Variational Stochastic Dirichlet Process-VAE, a variational autoencoder framework through replacing the standard variational distribution by a variational stochastic Dirichlet process (VSDP). Among our proposed framework, the encoder, leveraging stick-breaking constructions, adaptively partitions the latent space into clusters, while the decoder designed to recover each component achieves implicit signal separation. Its advantage is that the reconstruction target can be shifted from the raw input to its individual components. Experiments demonstrate our method’s efficacy in two scenarios: (1) Under coarse-grained source definitions, it reaches near-state-of-the-art performance (SDR=10.3); (2) For fine-grained track separation, what’s more, it identifies 83% of individual vocal tracks with an average SDR of 7.8, which cannot be obtained by other SOTA methods without the help of annotations.
Molecular activity prediction is a critical task in drug discovery and materials design, whose core challenge lies in accurately predicting the biological or chemical activity of molecules based on their structure and related features. Molecular activity is determined by complex multimodal physicochemical characteristics. Traditional activity prediction methods, often based on single-modal machine learning approaches, struggle to comprehensively capture the intrinsic relationships between multimodal molecular features and activity, limiting their predictive accuracy. To address these challenges, this paper proposes a Quantum multimodal fusion model for molecular activity prediction(QMF-MAP). QMF-MAP extracts and integrates multimodal features of molecules-including 1D fingerprint features, 2D graph structural features, and 3D spatial conformation features. These multimodal features are encoded into quantum states and fused via a quantum attention module. The fused multimodal representation is then processed by a quantum neural network to obtain discriminative results for activity prediction. The model was evaluated in classification comparison experiments on eight standard bioinformatics datasets. The results demonstrate that QMF-MAP achieves superior or comparable performance to the best traditional methods across multiple metrics, including classification AUC and accuracy. Further parameter analysis experiments confirm that QMF-MAP exhibits effective multimodal fusion capability and the proposed model demonstrates strong robustness against common near-term intermediate-scale quantum(NISQ) noise including amplitude damping, depolarizing, and bit/phase flip errors.
Anomaly detection (AD) aims to identify defects using normal-only training data. Existing anomaly detection benchmarks (e.g., MVTec-AD with 15 categories) cover only a narrow range of categories, limiting the evaluation of cross-context generalization and scalability. We introduce ADNet, a large-scale, multi-domain benchmark comprising 380 categories aggregated from 49 publicly available datasets across Electronics, Industry, Agrifood, Infrastructure, and Medical domains. The benchmark includes a total of 196,294 RGB images, consisting of 116,192 normal samples for training and 80,102 test images, of which 60,311 are anomalous. All images are standardized with MVTec-style pixel-level annotations, and anomalous samples are further enriched with structured text descriptions spanning both spatial and visual attributes, enabling multimodal anomaly detection tasks. Extensive experiments reveal a clear scalability challenge: existing state-of-the-art methods achieve 90.6% I-AUROC in one-for-one settings but drop to 78.5% when scaling to all 380 categories in a multi-class setting. To this end, we propose Dinomalym, a context-guided Mixture-of-Experts extension of Dinomaly that forms image-conditioned decoder feed-forward networks through convex combinations of expert parameter banks. It achieves 83.2% I-AUROC and 93.1% P-AUROC, outperforming existing approaches while maintaining measured latency close to vanilla Dinomaly. We aim to make ADNet a standardized and extensible benchmark, supporting the community in expanding anomaly detection datasets across diverse domains and providing a scalable foundation for AD foundation models. Dataset: https://grainnet.github.io/ADNet.
The present study provides the first affective norm for 3,783 Chinese words for ten discrete emotions: happiness, anger, sadness, fear, disgust, anxiety, surprise, contentment, amusement, and serenity. Ratings of words were collected from a total of 797 native speakers of Chinese using five-point Likert scales. A total of 895,858 valid ratings were collected. Results showed high inter-rater reliability and sufficient validity. This database shows an unequal distribution across the ten discrete emotions, with a total of 361 (9.54%) "neutral" words (words with an average rating of less than 3 in the ten discrete emotions), 2,482 (65.61%) "pure" words (words with an average rating of 3 or higher in only one discrete emotion), and 940 (24.85%) "mixed" words (words with an average rating of 3 or higher in more than one discrete emotion). The highest percentage of pure words was in the discrete emotion of serenity; a mixed word can be related to at most six different discrete emotions. Analyses revealed significant relationships between discrete emotions, affective dimensions (valence and arousal), and psycholinguistic variables (e.g., objective frequency, subjective frequency, familiarity, age of acquisition, concreteness, imageability, sensory experience rating, and emotional prototypicality). Small but significant individual differences in gender and personality were also revealed. To date, this new database is the largest set of ratings of ten discrete emotion categories in any language, and can serve as a valuable resource in studies of language, emotion, and cognition. The norms are available at https://osf.io/uztvh/?view_only=3d56e6dfcfb043eba2d3fb850ee27f68 .