ABSTRACT With the widespread deployment of object detection models based on deep neural networks (DNNs), their vulnerability to adversarial attacks in the physical domain has emerged as a significant security threat, stimulating substantial research interest within the academic community. This survey focusses on physical‐domain adversarial attacks targeting object detection models, presenting a comprehensive and in‐depth analysis of tasks including human detection, traffic sign detection and vehicle detection. Initially, we summarise the core concepts underpinning this research area. Subsequently, by examining the limitations of existing surveys, we delineate the core content and scope of this work. We then survey the latest research advances in this field and critically analyse the shortcomings of existing techniques. Finally, we outline promising future research directions and technical approaches. This narrative survey uses a transparent, PRISMA‐inspired search process to establish a solid foundation for researchers and practitioners to advance their comprehension of physical‐domain adversarial attacks.
ABSTRACT We present TDepth, a lightweight hybrid model for monocular depth estimation trained using teacher‐generated pseudo labels. TDepth combines a small CNN backbone, a local‐global transformer encoder, an attention‐guided skip‐fusion decoder, a bin‐based depth head and a lightweight refiner for improving boundaries and thin structures. To avoid dependence on expensive sensor depth during training, we generate pseudo depth labels offline using Pixel‐Perfect‐Depth and train on about 550K images from NYU Depth v2 and ScanNet++. We evaluate TDepth on in‐domain, zero‐shot and efficiency settings. On NYU Depth v2, TDepth‐Full achieves , AbsRel and RMSE , while the smaller 15.1M‐parameter variant remains competitive with much lower model complexity. In zero‐shot evaluation, TDepth‐Full achieves and AbsRel on iBims‐1 and and AbsRel on SUN‐RGBD, showing strong transfer to unseen indoor datasets. TDepth‐Small reaches about 240 FPS at and remains above 30 FPS even at , while TDepth‐Full reaches 60 FPS at and provides stronger depth accuracy with a practical runtime profile. These results show that TDepth provides a strong balance between depth accuracy, zero‐shot generalisation and computational efficiency, making it a practical option for lightweight monocular depth estimation.
ABSTRACT Multi‐modal reasoning enhanced by large language models is central to developing vision‐language models (VLMs). However, existing methods ignore extracting local details with global context when handling complex vision‐language tasks. This study introduces a novel In‐Context Learning into Chain‐of‐Thought reasoning (IC2T) framework, adopting a staged reasoning mechanism that progressively optimises the integration of local and global features to satisfy the demands of complex tasks. The IC2T dynamically extracts the regions of interest as the local information and the entire vision as the global information, which are then fused and stepped into multi‐round reasoning. The proposed in‐context learning scheme retrieves similar samples and uses their global and local information to provide contextual guidance for progressive reasoning. The extensive experimental results demonstrate that IC2T based on multi‐scale VLMs outperforms state‐of‐the‐art models across multiple multi‐modal benchmarks, and extensive ablation studies highlight the critical roles of the proposed in‐context learning and chain‐of‐thought reasoning in improving logical reasoning performance.
ABSTRACT The proposed G‐AGF Net is a radar‐camera fusion 3D object detection framework characterised by physical awareness and environmental adaptability. Specifically designed to address geometric degradation caused by visual perception deterioration in challenging environments, the model utilises Anisotropic Gaussian Splatting (AGS) to reshape the physical distribution of sparse point clouds using radar Doppler velocity. Furthermore, an Environment‐Aware Module (EAM) is integrated to quantify visual reliability via the environmental reliability score in real‐time. Finally, the Gated View Transformer (GVT) leverages this score to anchor image features using the radar‐derived probabilistic depth prior. Experimental results on the nuScenes dataset demonstrate that G‐AGF Net achieves an NDS of 0.702 across all scenarios and exhibits superior robustness in the evaluated night and rainy conditions, demonstrating its effectiveness under the challenging environmental conditions considered in this study.
ABSTRACT Convolutional Neural Networks (CNNs) and Transformers have achieved remarkable success in hyperspectral image (HSI) classification. However, CNNs are limited in modelling long‐range spatial dependencies, while the quadratic computational complexity of Transformer restricts their application to high‐dimensional HSI data. Recently, state‐space models (SSMs), particularly Mamba, have shown great potential due to their linear computational complexity and powerful sequence modelling capabilities. However, the original Mamba model is designed for one‐dimensional sequence modelling and struggles to effectively handle the inherent three‐dimensional spatial‐spectral characteristics of HSI. Furthermore, it lacks dedicated considerations for the deep fusion of spectral and spatial information and the integration of global context and local details. To address this issue, we propose a novel Multi‐Scale Spatial‐Spectral Mamba Network (MSS‐MambaNet) for efficient and accurate HSI classification. Specifically, MSS‐MambaNet consists of four key components: First, the Multi‐scale Local Spectral Feature Extraction Module (MSLSFE) captures spatial‐spectral features at different scales through a parallel multi‐branch architecture. Second, the Spatial Mamba Encoder employs an innovative dimensionality reorganization strategy to model global spatial context while preserving spatial structure. Third, the Spectral Mamba Encoder incorporates a bidirectional split‐scan mechanism to comprehensively capture the complex dependencies between spectral bands. Finally, the Global Spatial‐Spectral Feature Fusion Module (GSSFF) deeply integrates the complementary information of spatial and spectral features through multi‐scale dilated convolutions and a detail‐enhanced attention mechanism. MSS‐MambaNet achieves overall accuracies of 97.19%, 94.25%, and 97.38% on the Pavia University, Indian Pines, and Salinas datasets, respectively, outperforming state‐of‐the‐art methods in both classification accuracy and computational efficiency. The code will be available at https://github.com/ZhangSaida/Mss‐MambaNet .
