Efficient deployment of Vision Transformers is increasingly important for real-time visual computing. Although 4/8-bit ViTs often preserve higher accuracy than binary ViTs, their dense projection layers still limit realized acceleration. We propose LiteBDS-ViT, a hardware-friendly structural optimization framework that replaces dense attention and feed-forward projections with block-diagonal shift projections, allocates group numbers layer-wise under a global operation budget, and implements fused CUDA kernels to translate theoretical sparsity into practical GPU speedup. Across small-scale classification, ImageNet-1K large-scale classification, and ADE20K semantic segmentation, LiteBDS-ViT improves the accuracy–latency trade-off over representative binary and quantized ViTs, achieving up to 1.45× model forward inference speedup under comparable operation budgets. The released repository, including source code, pretrained models, training scripts, inference kernels, and configuration files for reproducing our experiments, is publicly available at https://github.com/WanRui37/LiteBDS-ViT . It is also permanently archived at https://doi.org/10.5281/zenodo.19437156 .
Autonomous navigation in complex, dynamic environments remains a fundamental challenge for unmanned aerial vehicles (UAVs), requiring robust perception, decision-making, and control capabilities. While conventional planning methods lack real-time adaptability, deep reinforcement learning approaches often demand excessive interaction data with high trial-and-error costs. This paper presents a data-efficient imitation learning framework that combines generative adversarial imitation learning (GAIL) with a sequential trajectory inpainting technique to learn navigation policies directly from demonstration data. The framework uses an eight-dimensional state representation (3D position, velocity, yaw, and pitch) and a five-dimensional continuous action space. Given only 10 expert trajectories (each containing 10,000 state-action pairs) collected from a human expert operator in a high-fidelity AirSim simulation environment, our method leverages a GAN architecture enhanced with long short-term memory (LSTM) networks to repair and augment incomplete expert trajectories, ensuring both distributional and temporal consistency. The navigation policy is then trained through GAIL utilizing proximal policy optimization (PPO) for stable policy updates. Extensive simulations in dynamic pursuit and restricted-zone penetration scenarios validate our approach against multiple baselines including DDPG, PPO, APF, and ACO. Our method achieves an 87.88 https://github.com/yuning715/GAIL-UAV-Path-planning.git.
Video anomaly detection (VAD) is crucial for intelligent surveillance and public safety, identifying events that deviate from normal patterns without anomaly samples. This paper introduces a VAD method leveraging a multi-source pseudo-anomaly generator and a dual-decoder framework. Our approach integrates spatiotemporal perturbations and object-level local appearance perturbations to enhance anomaly coverage. A dual-path decoder architecture, equipped with independent memory modules, encourages the separation of normal and pseudo-anomalous representations in the latent space, discouraging accurate reconstruction of pseudo-anomalous samples and increasing their reconstruction discrepancy from normal samples. Experiments on UCSD Ped2, CUHK Avenue, and ShanghaiTech yielded frame-level AUC values of 98.7 https://github.com/SVIL2024/DualPath-PseudoAnomaly-VAD .
Domain Generalization aims to develop models capable of learning from one or several related source domains and generalizing to unknown target domains, addressing challenges posed by distributional shifts. Existing methods often rely on fixed causal relationships, limiting their adaptability during deep network training. To address this gap, we propose dynamic spectral-causal risk minimization (DSCM) for domain generalization. DSCM introduces two key components, namely a dynamic causal masking mechanism that adaptively suppresses spurious features while retaining causal ones, and a spectral causal consistency (SCC) loss that aligns gradient spectral energy across sub-environments to enhance cross-domain stability. Experimental results on PACS, VLCS, OfficeHome, TerraInc, and DomainNet datasets demonstrate competitive performance with DSCM achieving average accuracies of 86.7, 77.7, 70.9, 48.6, and 42.9 https://github.com/xtcjswxwx/DSCM.
Image segmentation is a core visual computing task that requires both semantic consistency and boundary fidelity in visually degraded environments. In specialized settings such as clinical endoscopy, underwater scenes, and camouflage, low contrast, small targets, fuzzy boundaries, and domain shifts make this requirement difficult. We introduce DPSAM2, a frozen-backbone adaptation framework for visually challenging segmentation tasks. DPSAM2 integrates a Multi-Field Bottleneck Fusion (MFB) module to combine complementary local-detail and contextual cues, a Dynamic Memory Bank (DMB) that maintains compact cross-sample structural references through quality-aware training-time updates, and a SAM2-anchored residual dual-path decoder that applies bounded detail correction to SAM2 semantic logits. All evaluations are conducted in a strictly prompt-free setting at test time, enabling direct comparison with conventional automatic segmentation models. Across 10 benchmarks spanning polyp, marine animal, and camouflaged object segmentation, DPSAM2 achieves competitive overall performance, with representative gains of +5.4 percentage points in mDice on ETIS, +3.1 percentage points in mIoU on RMAS, and +3.8 percentage points in F^w_β on CHAMELEON. These results are consistent with DPSAM2 being a practical framework for visual segmentation scenarios where small targets, fuzzy boundaries, and domain shifts degrade dense predictions. To support reproducible open science, we provide the project repository at https://github.com/wxshCoding/mmsam2.git .
Reliable localization of partially visible humans is an essential front-end requirement for vision-based drowning-risk monitoring. However, horizontal-view water-surface surveillance remains challenging because visible human regions are often small and incomplete, while waves, foam, reflections, motion blur, and lens contamination introduce strong background interference. This paper presents AHD-YOLO, a lightweight detector based on YOLO11n for horizontal-view water-surface human-part detection. The proposed framework does not perform single-frame drowning-event recognition; instead, it detects exposed heads and hands to provide spatial observations for subsequent tracking, temporal behavior analysis, and alarm decision modules. A layer-wise hybrid downsampling strategy combines robust feature downsampling (DRFD) and Haar wavelet downsampling (HWD) to preserve salient semantic responses and high-frequency boundary details at different feature levels. In addition, an asymmetric cross-domain attention (ACA) module recalibrates multi-scale fused features to enhance target-related responses and suppress water-surface interference. A dedicated dataset containing 1,222 source images is reannotated with head and hand bounding boxes and expanded to 6,110 images using reflection, water stain, camera shake, and mixed degradations. Experimental results show that AHD-YOLO achieves 92.47 mAP_50 and 57.18 mAP_50–95 with 3.06M parameters and 3.5 GFLOPs. These results demonstrate that the proposed model provides an effective accuracy–complexity trade-off for front-end perception in real-time drowning-risk monitoring.
Vision-based 6D pose estimation is critical for augmented reality-assisted assembly, human–robot collaboration, and quality inspection in intelligent manufacturing. However, performance degrades severely in complex industrial scenarios due to occlusion, varying lighting, textureless surfaces, and reflective parts. This work presents a monocular 6D pose estimation approach using hybrid representations and adaptive multi-feature fusion to address these challenges. A hybrid representation learning framework is designed to jointly predict keypoint heatmaps, relational vectors, semantic edges, masks, and visibility, thereby enhancing feature robustness. A multi-feature adaptive fusion strategy optimizes the pose by combining semantic and fine-grained general features. A structure-constrained correction module refines multi-object poses using assembly consistency constraints. Experiments on a custom industrial assembly dataset and the public Mono6D dataset show that the proposed method achieves 87.46 ^∘ accuracy, outperforming state-of-the-art methods. The custom dataset includes multiple weakly textured and reflective assembly parts under occlusion, lighting variation, and multi-viewpoint conditions. Furthermore, the complete system runs at approximately 18 FPS, with faster tracking once initialized. The approach supports reliable AR-assisted assembly and meets industrial deployment requirements. Our code and datasets are open-sourced at https://github.com/nengbinlv/HRMFPose , with the DOI: https://doi.org/ https://doi.org/10.5281/zenodo.19574143 .
Photomosaic systems have long treated the tile library as a purely visual dataset, selecting tiles by color and edge similarity alone. This paper presents a rendering architecture in which an emoji tile library is treated instead as a semantically annotated dataset, governed by a five-signal scoring model operating over an adaptively subdivided diamond tile grid. The architecture introduces a three-geometry artist region system, a two-pass scoring model, a semantic theme scoring layer that biases candidate selection toward emoji whose annotations match the source image’s theme vocabulary without overriding perceptual quality, and an iterative palette correction mechanism that corrects systematic candidate routing errors using render-history feedback. The architecture is demonstrated across five works whose differing painting techniques each motivate a distinct architectural response. Controlled comparisons from per-tile placement logs show that semantic theme scoring raises on-theme placement rate on each of the four works with theme presets configured, by 6.6–13.6 percentage points depending on the work’s palette and theme vocabulary, with no meaningful change in placement entropy. A separate comparison shows that palette correction reduces systematic lightness and chroma residuals by 40
Vision transformer models have achieved remarkable success in computer vision, with Swin transformer serving as a versatile hierarchical backbone for classification, detection, and segmentation tasks. However, increasing model scales bring prohibitive memory consumption and computational costs that hinder deployment in resource-constrained environments. This paper proposes MoR-ST, a content-adaptive recursive computation framework that integrates the mixture of recursions (MoR) paradigm into the Swin transformer architecture. The core mechanism is an adaptive token-level recursion strategy: a lightweight routing network dynamically assigns recursion depths to individual tokens based on their semantic complexity, enabling simple tokens (e.g., background regions) to exit early while complex tokens (e.g., textured objects) undergo deeper processing. We evaluate MoR-ST on three standard benchmarks: ImageNet-1K/22K image classification, COCO object detection, and ADE20K semantic segmentation. Experimental results show that MoR-ST achieves competitive accuracy compared to the original Swin transformer while reducing parameters by approximately 50 × . On ImageNet-1K, MoR-Swin-T attains 80.5 ∼ 50 × speedup, and the value of this trade-off depends on the deployment regime, which we quantify rather than assert. These results demonstrate that adaptive recursive computation provides an effective pathway for building efficient vision backbones suitable for real-time visual computing applications.
3D shape recognition is pivotal in computer vision, with applications spanning autonomous driving, digital twins, and cultural heritage preservation. This paper introduces MVRT-Net, a novel Multi-View Relational Transformer Network, addressing the challenge of balancing local details and global context in 3D shape descriptors. MVRT-Net integrates a Multi-Scale Adaptive Dynamic Graph Attention mechanism and a Transformer-based global view relational modeling strategy with adaptive key-view selection and adaptive attention pooling, which dynamically learns view-level importance weights to suppress redundant information and fuse cross-scale semantics into a more discriminative global descriptor. Experimental results on the ModelNet40 dataset demonstrate MVRT-Net’s superiority, achieving a classification accuracy of 97.89 https://github.com/weihai0226/MVRT/tree/master/MVRT .
Mesh inpainting remains a critical challenge in geometry processing, particularly for real-world scans that often contain holes and missing regions. This paper introduces a data-free neural mesh inpainting framework that operates directly on incomplete input meshes while preserving the original mesh connectivity. Our method first constructs an intermediate watertight surface from a zero-level isosurface, where projected hole boundaries guide geodesic patch cutting and stitching. A self-supervised graph convolutional network then refines the repaired regions. Experiments demonstrate that our method outperforms traditional dataset-independent approaches and exhibits greater robustness than existing deep learning models. The repaired meshes maintain sufficient integrity for downstream applications such as physical simulation and 3D printing. Code is available at https://github.com/ZHLRJ/Mesh-Inpainting .
Adversarial attacks against embodied navigation agents have been demonstrated to significantly disrupt navigation direction and trajectory. However, existing adversarial attack methods for embodied navigation face two major challenges: (1) They primarily rely on 2D pixel perturbations within a limited set of viewpoints, struggling to mislead the agent under unseen viewing angles. (2) These methods optimize perturbations under static viewpoints, leading to attack failures in real-world scenarios with dynamic viewpoint changes. To overcome these limitations, we propose a novel adversarial attacks via 3D adversarial examples (AA3D) for embodied object navigation to realize the efficient deception of the agent in dynamic scenes. Specifically, our AA3D incorporates two core designs: radius-adaptive view modulation (RAVM) and dual-sampling optimization strategy (DSOS). RAVM adaptively modulates viewing parameters of camera to capture abundant nonstandard viewpoints, thus enhancing the attack performance of 3D adversarial objects across a wide range of viewpoints. DSOS utilizes differentiable rendering to optimize adversarial textures based on stochastic and trajectory-aware viewpoints, ensuring consistent deception of the agent during continuous and dynamically changing viewpoint transitions. Benefiting from the above designs, our AA3D persistently disrupts the navigation trajectory of the agent and misleads its decision-making in variable navigation scenarios, even under extreme viewpoints. Extensive quantitative evaluations on the Habitat-Matterport3D (HM3D) dataset demonstrate the adversarial effectiveness of our method. Compared with the baseline methods, our AA3D reduces the SR to 50 https://github.com/Delancey00/3D-attack-in-nav .
The Bidirectional Reflectance Distribution Function (BRDF) is widely used in computer graphics to model material appearance based on light and view directions, though it neglects other dependencies, such as spatial. While numerous parametric BRDF models exist, they vary significantly in complexity and their capacity to fit empirical data. This paper evaluates the automated performance of 21 frequently used BRDF models, focusing on their utility without human tuning. We utilized the Levenberg–Marquardt method for parameter estimation without prior knowledge or post-processing. Our motivation stems from the need for efficient modeling more complex and accurate approximations of material visual appearance in the form of Spatially Varying BRDF (SVBRDF) and Bidirectional Texture Functions (BTF). While SVBRDFs offer lower complexity than full BTFs, both can be effectively combined. Our goal was to compare these models using a fully automatic approach to determine their suitability for SVBRDF or BTF modeling, particularly with restricted angular resolution measurements for both isotropic and anisotropic materials. To ensure accuracy, all tests were conducted using high-quality UTIA BTF measurements (from the MAM2014 sample set). Factual errors were computed directly between measurements and model predictions to eliminate potential contamination from rendering or illumination inaccuracies. Ultimately, we aim to identify the best automatically estimated general model capable of representing a wide variety of materials.
Generating complete Chinese font libraries from limited reference characters is challenging due to complex character hierarchies and diverse style variations. Few-shot font generation approaches often fail to preserve structural integrity or achieve high style fidelity across unseen fonts. Here, we propose IM-Font, a one-shot font generation method that leverages Ideographic Description Sequence (IDS) structural encoding and a multi-stage style injection mechanism within a diffusion framework. Unlike existing diffusion-based or static IDS-aware methods, we introduce a timestep-partitioned injection mechanism that aligns hierarchical IDS encoding with distinct diffusion denoising stages, enabling a progressive transition from coarse structural guiding to fine-grained style rendering. IDS encoding explicitly models the hierarchical composition of characters, ensuring stroke completeness and component accuracy. The multi-stage style injection incorporates specific style features at designated denoising stages, enabling fine-grained texture control while maintaining structural coherence. Experiments on a benchmark of 100 font styles show that IM-Font outperforms six state-of-the-art methods, achieving a 39.45 https://github.com/ZhangRan-lucky/IM-Font/ and permanently archived at https://doi.org/10.5281/zenodo.19422461 .
Brain tumor segmentation under partial modality Magnetic Resonance Imaging (MRI) scenarios remains challenging because segmentation models may exhibit strong modality dependence and prediction instability when the available inputs change from complete multimodal MRI to a single modality or a reduced modality subset. To address this problem, this paper proposes OW-Mamba, a modality adaptive 3D segmentation framework for accurate and stable single modality inference while supporting arbitrary valid modality combinations. The methodological idea is to encode modality availability and feature level confidence modulation, learn a shared structural representation that adapts to different inputs, and preserve complementary appearance cues when multiple modalities are available. OW-Mamba employs a backbone based on state space modeling to capture long range spatial dependencies and preserve global tumor structure under limited modality information. It further introduces a modality adaptive base residual fusion strategy for stable feature integration with controlled local refinement. A feature level confidence modulation mechanism suppresses low confidence voxel level responses, and bottleneck level channel calibration mitigates distribution shifts across available modalities. Experiments on two standardized public brain tumor MRI benchmarks under single-modality and partial-modality settings provide benchmark-level methodological evidence of stable and accurate segmentation across MRI inputs. Across all 15 modality combinations, OW-Mamba achieves average Dice Similarity Coefficient values of 84.5 https://github.com/Curmudgeon-zhou/OW-Mamba-public .
Simultaneous Localization and Mapping (SLAM) is central to robotic perception because it enables robots to estimate their motion while building maps of unfamiliar environments. Recent 3D Gaussian Splatting (3DGS)-based SLAM methods can produce dense and renderable scene representations, but the continuous accumulation of Gaussian primitives increases the cost of online rendering and optimization. This paper presents MS-SLAM, a memory-guided 3DGS-SLAM framework for efficient robotic deployment. MS-SLAM maintains a long-term Gaussian map while constructing a short-term tracking map and a pose-dependent recall map, so only task-relevant Gaussian subsets participate in pose estimation and map optimization. Experiments on Replica, TUM-RGBD, ScanNet, ETH3D, a robotic arm, and a mobile robot show that MS-SLAM stabilizes per-frame processing while maintaining competitive trajectory accuracy and rendering quality. On Replica/R0 and TUM/fr1d, MS-SLAM reduces per-frame tracking time compared with representative Gaussian SLAM baselines while preserving comparable ATE RMSE. Robot experiments further demonstrate the feasibility of online RGB-D mapping in real indoor scenes. Code is available at https://github.com/785229637/MS-SLAM .
Facial attribute editing (FAE) enables realistic facial property manipulation but also raises risks of malicious forgery. Existing forensic methods suffer from poor generalization across editing techniques and vulnerability to post-processing like JPEG compression. This work introduces the Boundary-Guided Frequency Fusion framework for unified image-level detection and pixel-level localization of FAE. The RGB branch provides stable semantic localization, while a DCT-based high-frequency branch offers complementary residual cues. A learnable gated fusion mechanism suppresses spurious noise, and boundary supervision sharpens localization edges. Experiments on CelebA-HQ benchmarks show that the method achieves 99.77 https://github.com/oceanpromax/gjw .
Few-shot 3D point-cloud semantic segmentation aims to recognize novel classes from limited annotated support samples, but current prototype-based methods often match query and support points in a shared feature space. This coupling can mix geometric structure with appearance cues such as color, intensity, and local density, which weakens matching when query scenes contain visually similar distractors. We propose D2M-Net, a decoupled dual-matching network that separates backbone features into geometry-oriented and semantic-oriented subspaces before prototype comparison. Two lightweight branches with LayerNorm produce the decoupled representations, and an orthogonal regularization term encourages the branches to encode complementary information. D2M-Net then computes prototype similarities in the two subspaces and uses a point-wise adaptive fusion gate to weight the more reliable cue for each query point. A dual-path background calibration module further suppresses false activations from base-class regions. Experiments on S3DIS and ScanNet show consistent gains over recent few-shot 3D segmentation baselines. On S3DIS, D2M-Net improves the 1-way 1-shot mean intersection-over-union (mIoU) from 47.21 to 50.34
Low-light image enhancement aims to recover visually reliable images from underexposed scenes while preserving color, structure, and robustness to noise. Existing unsupervised Retinex-based methods often estimate reflectance and illumination from spatial features in which illumination attenuation, color degradation, texture loss, and noise responses are coupled. We propose SF-GAL, a prior-calibrated spatial-frequency Retinex decomposition framework for unsupervised low-light image enhancement. SF-GAL calibrates a CLAHE-derived structural prior, decomposes low-light features through complementary spatial and wavelet branches, and uses structure-guided frequency modulation (SGFM) to regulate frequency responses before Retinex prediction. The enhanced image is reconstructed through adaptive Retinex recomposition with guided filtering and exposure adjustment. Experiments on LOL, SICE, and LSRW show competitive unsupervised performance with dataset- and metric-dependent strengths. Among the unsupervised methods, SF-GAL obtains the lowest Δ E on SICE and the best PSNR, MAE, and Δ E on LSRW (Huawei), while competing methods lead the remaining metrics and subsets. Ablation studies, mechanism visualizations, camera-specific analyses, and nighttime semantic segmentation evaluation support the complementary roles of SPC, wavelet decomposition, and SGFM. The source code and split files are publicly available at https://github.com/airwaves0/sfgal .
Weakly supervised semantic segmentation (WSSS) based on image-level labels typically relies on class activation maps (CAMs) to generate pseudo-labels. However, single-layer prototype-based generation methods often result in incomplete CAM activation and inaccurate target localization due to limited feature representation capabilities. Furthermore, the reliance of existing methods on discriminative appearance information makes it difficult to effectively utilize object structural information, further impacting boundary accuracy and the final segmentation performance. To address this problem, this paper proposes a complementary prototype representation framework, employing three modules to collaboratively improve pseudo-label quality. First, the Multi-scale prototype fusion module (MSPF) adaptively integrates multi-level features to construct a more complete and accurate category representation. Second, the structural–appearance prototype decoupling module (SAPD) separates appearance and geometric information in pixel-prototype matching, enhancing texture discrimination and boundary details. Finally, Negative prototype clustering mining (NPCM) improves the discriminative ability of confused categories by generating semantically similar sub-category negative samples. Experiments on the PASCAL VOC 2012 and MS COCO 2014 datasets demonstrate that our method achieves 73.3 https://github.com/Weiw1819/MSPD-Net .
