Partially overlapping point cloud registration is fundamental in computer vision and robotics, yet remains challenging due to inaccurate overlap identification and unreliable correspondence extraction. Existing learning-based methods often apply indiscriminate positional encoding to all points and insufficiently exploit intrinsic geometric structures, which limits registration precision under partial overlap. To address these issues, we propose OOGM, an overlap-oriented geometric matching method that localizes positional encoding within high-confidence overlapping regions and explicitly embeds transformation-invariant geometric structures. OOGM first constructs compact geometric descriptors to provide explicit local geometric context, then develops an overlap-aware matching strategy to restrict feature aggregation and structural modeling to promising overlapping areas, reducing interference from non-overlapping points. A consistency-driven mechanism further evaluates correspondences through hierarchical geometric consistency and calibrates confidence scores for reliable match selection. Extensive experiments on synthetic and real-world datasets demonstrate the effectiveness of OOGM, with 26% and 45% reductions in rotation and translation errors on ModelNet40 compared with competitive baselines.
Implicit Neural Representations (INRs) represent a new paradigm in signal representation, have recently achieved success in various visual tasks. To address the low-frequency bias issue of Multi-Layer Perceptrons (MLPs), researchers have conducted extensive works on activation functions, positional encoding and network structures. However, existing INR methods generally adopt the uniform sampling strategy during the learning process, ignoring the imbalance in learning efficiency among different-frequency components. To tackle this problem, we introduce a residual-aware training (RAT) scheme to enhance INR models. RAT constructs a dynamic residual map through exponential moving average, which reflects the learning efficiency of different-frequency components. Based on this map, RAT adaptively adjusts the sampling frequency and loss function, enhancing the learning ability of MLPs. We validate the effectiveness of RAT scheme on different vision tasks with various representative INR models, experimental results show that armed with our RAT scheme, mainstream INR models could improve reconstruction fidelity in terms of PSNR and visual quality. The codes will be released at https://github.com/Yangfan-123-cell/RAINR.
While Dynamic 3D Gaussian Splatting enables real-time novel view synthesis of dynamic scenes, its prohibitive memory requirements severely limit practical applications, often demanding several gigabytes for even short sequences. To address this limitation, we introduce MACGS, a comprehensive compression framework that significantly reduces memory requirements while preserving rendering quality. At its core, MACGS employs a motion-aware pruning strategy that differentially processes dynamic and static regions, coupled with a spatial-aware compression scheme that selectively reduces model parameters while preserving temporal modeling capabilities. Our framework integrates resolution-adaptive redundancy analysis with asymmetric parameter compression, enabling efficient optimization of both Gaussian primitives representation and network structure. Compared with 4DGS, MACGS achieves a 338× reduction in storage requirements (from 3.1GB to 9.24MB) while maintaining visual quality with less than 0.6dB PSNR drop and enabling real-time rendering at over 130 FPS. Our method achieves a compelling balance between extreme compression and fidelity, significantly outperforming existing compression-focused approaches by reducing storage requirements by 2.7× compared to the current state-of-the-art methods while maintaining comparable visual quality. Code is available at https://anonymous.4open.science/r/MACGS.
The fusion of hyperspectral (HS) and Light Detection and Ranging (LiDAR) data plays a crucial role in enhancing land-cover classification by jointly exploiting spectral, spatial, and structural cues. However, existing multimodal fusion methods still struggle to model long-range dependencies and complex anisotropic interactions while maintaining computational efficiency. This paper introduces M2Heat, a physics-inspired framework that investigates multimodal fusion through the lens of heat conduction. At its core, a physics-driven visual heat conduction module (vHeat) and enhanced Frequency Value Embeddings (FVEs) simulate anisotropic information flow, enabling the capture of global dependencies with sub-quadratic complexity and physical interpretability. This mechanism, combined with a hybrid spatial-frequency fusion strategy named Cross-Frequency Fusion (CFF) module, produces highly discriminative and robust feature representations. M2Heat achieves competitive overall performance on three benchmarks, i.e., Trento, Houston2013, and Augsburg, while providing an interpretable heat-conduction-guided perspective for multimodal feature fusion. These results indicate the potential of heat-conduction-guided neural operators for efficient and interpretable RS multimodal fusion. The source code is publicly available at https: /github.com/Weikan0425/M2Heat_HSI_LiDAR.
SSelective light attenuation and frequency-dependent degradation are common features of aquatic environments. Because of these physical conditions, underwater image enhancement (UIE) is highly challenging. Recent deep learning techniques have done very well at enhancing visual contrast. But most deep learning techniques use data-driven models that employ black box techniques. When tested in different water bodies, they often produce unrealistic artifacts. Traditional physical models can ensure color consistency. Their rigid priors, however, destroy fine high-frequency details. We build a Dual-Stream Wavelet-PCN framework to break this bottleneck. It firmly binds frequency-domain analysis with physics-constrained learning. A parallel Wavelet Stream uses the discrete wavelet transform (DWT). It completely strips high-frequency textures from the low-frequency background structure. This mechanism directly isolates the physical haze from edge details. Simultaneously, we introduce a Physics-Constrained Network (PCN) Stream to strictly constrain color restoration based on the underwater image formation model. To overcome uniform lighting assumptions and hardcoded empirical hyperparameters, we introduce a Dual-Branch Backlight Estimation strategy, employing end-to-end learnable parameters to dynamically fuse global and local illumination cues for non-uniform light fields. Finally, an Adaptive Attention Fusion module dynamically integrates complementary features from both streams. Extensive experiments on four benchmarks show that our method achieves strong structural visibility and competitive performance on several non-reference perceptual metrics, indicating a stable balance between color correction, detail recovery, and artifact suppression. Code is available at https://github.com/ubpinn/DS-WPCN.
