voxel grid processing

This work addresses the limitations of existing single-image 3D scene generation methods, which often rely on low-resolution representations and struggle to reconstruct high-fidelity, complete scenes. To overcome this, the authors propose a hierarchical voxel-enhanced framework for high-fidelity 3D scene synthesis. Given an input image, the method first segments it and leverages an attention mechanism to retrieve and align corresponding 2D–3D components, constructing a hierarchical component tree. It then performs coarse-to-fine voxel super-resolution reconstruction, generating high-resolution 3D scenes while preserving both geometric and semantic consistency at the component level. Extensive evaluations demonstrate that the proposed approach significantly outperforms current state-of-the-art methods across multiple metrics, achieving superior performance in 3D scene generation quality.

To address low memory efficiency, high computational overhead, and limited GPU support in real-time, resource-constrained 3D surface reconstruction, this paper proposes a variance-adaptive multi-resolution voxel grid. It dynamically adjusts voxel size based on local Signed Distance Function (SDF) observation variance and employs a flattened spatial hash for constant-time memory access and full GPU parallelism. Unlike conventional octrees, it integrates a parallel quadtree to regulate Gaussian lattice density. The key innovation lies in the first tight integration of variance-driven adaptive voxel partitioning with hash-based indexing, significantly improving scalability and rendering efficiency. Our open-source implementation achieves up to 13× speedup and 4× memory reduction over fixed-resolution baselines, while maintaining comparable reconstruction accuracy—enabling real-time, high-performance 3D reconstruction and rendering.

Addressing the challenge of simultaneously achieving high geometric accuracy, completeness, and fine-detail fidelity in sparse-voxel surface reconstruction, this paper proposes GeoSVR—an explicit volumetric framework. Our method integrates differentiable rendering with explicit voxel optimization while introducing two key innovations: (1) uncertainty-aware depth supervision, which leverages per-pixel uncertainty maps from monocular depth estimation to stabilize optimization and improve convergence robustness; and (2) sparse-voxel surface regularization, enforcing geometric consistency to enhance surface sharpness and topological integrity—especially for small-scale voxels. By jointly optimizing voxel occupancy and surface geometry under these constraints, GeoSVR achieves efficient computation without sacrificing reconstruction quality. Extensive experiments demonstrate that GeoSVR outperforms state-of-the-art radiance fields and both implicit and explicit reconstruction methods across diverse complex scenes, delivering superior geometric accuracy, more complete surface coverage, and better preservation of fine geometric details.

To address the limited receptive field and significant performance degradation at high bit depths in voxelized point cloud geometry compression, this paper proposes a staged spatial-to-channel (S2C) context modeling framework. The method introduces three key innovations: (1) a novel staged channel-wise autoregressive model that hierarchically captures coarse-grained neighborhood dependencies; (2) integration of spherical coordinate representation with geometric residual coding (GRC) to ensure consistent resolution across hierarchical predictions; and (3) a large-kernel residual probability approximation (RPA) module to enhance entropy estimation accuracy. Experimental results demonstrate that the proposed approach achieves substantial bitrate reduction while maintaining or even improving reconstruction quality. Moreover, its computational complexity is lower than that of state-of-the-art voxelized methods, with particularly pronounced advantages on dense point clouds encoded at high bit depths.

To address challenges in unordered point cloud geometry compression—namely, difficulty in modeling neighborhood relationships and quantization distortion introduced by voxelization—this paper proposes a context-adaptive residual coding framework based on KNN neighborhood modeling. The method employs a two-tier encoder-decoder architecture: a non-learned base tier efficiently reconstructs global structure, while an INR-driven learnable refinement tier models local geometric context to recover fine details and enable arbitrary-density upsampling at the decoder. Key contributions include: (i) the first incorporation of content-aware local geometric context into raw point cloud compression; (ii) a low-complexity two-tier architectural design; and (iii) the first integration of implicit neural representations (INRs) into a compression system that jointly supports compression and multi-scale reconstruction. Experiments demonstrate that our method reduces encoding latency and model complexity by two orders of magnitude over SOTA methods, while maintaining superior rate-distortion performance and enabling flexible surface sampling.

Existing city-scale 3D reconstruction methods struggle to produce complete, watertight, and geometrically regular meshes suitable for simulation. To address this limitation, this work proposes a scalable end-to-end image-to-mesh reconstruction framework that adopts a divide-and-conquer strategy: it avoids global feature matching through topological image clustering and spatial partitioning, enabling distributed reconstruction via block-wise sparse structure-from-motion (SfM) and geometry-aware camera selection. Furthermore, the method introduces a curvature-adaptive remeshing technique that dynamically adjusts vertex density to preserve fine details while enhancing geometric regularity. Notably, the approach operates without global optimization, efficiently generating high-fidelity, watertight, detail-rich, and topologically consistent 3D meshes across city-scale scenes.

This work proposes an end-to-end learned surface primitive compression framework to address the inefficiency of point cloud geometry compression and the lack of smoothness in reconstructed surfaces. The method introduces probabilistic surface elements (pSurfels) to model local point occupancy and constructs an adaptive octree structure, termed pSurfelTree, which incorporates a Tree Decision module for rate-distortion-optimized granularity selection. By leveraging a probabilistic generalized Gaussian distribution and a hierarchical representation, the framework effectively avoids redundant encoding in smooth regions, thereby achieving both high compression efficiency and significantly improved surface smoothness and coherence in reconstruction. Under MPEG common test conditions, the proposed approach outperforms voxel-based baselines and the G-PCC-GesTM-TriSoup standard in geometric compression performance.

This work addresses the challenge of degraded 3D surface reconstruction accuracy caused by missing geometric information in LiDAR point clouds due to limited scanning range and occlusions. To tackle this issue, the authors propose a reconstruction method based on plane classification and priority-driven growth. The approach categorizes scene planes into three visibility classes—highly visible, partially visible, and invisible—and employs a hierarchical spatial partitioning scheme. Coupled with a min-cut optimization strategy, it generates compact, watertight polygonal models that effectively recover missing geometric details. Evaluated on public datasets, the method significantly outperforms current state-of-the-art techniques, achieving higher reconstruction fidelity while preserving model compactness.

This paper presents a systematic review of learning-based 3D representations for tasks such as 3D reconstruction, novel view synthesis, and rendering, spanning from traditional explicit formulations—including meshes, point clouds, and voxels—to emerging implicit neural fields and primitive-based splatting techniques like 3D Gaussian Splatting. Emphasizing the evolutionary trajectory of 3D representations themselves, the work highlights the paradigm shift from explicit to implicit modeling, clarifying the mathematical formulations, strengths, limitations, and suitable application scenarios of each approach. Distinct from prior task-centric surveys, this study centers on representation as the core organizing principle, offering fresh insights for 3D/4D content generation and identifying key challenges and future research directions to serve as a theoretical reference for the computer graphics and vision communities.

The field of 3D vision suffers from fragmented data representations, learning paradigms, and benchmarking protocols, leading to a lack of unified understanding regarding efficiency, fidelity, and scalability. This work proposes the first cohesive conceptual framework that integrates geometric representations—such as point clouds, meshes, voxels, and 3D Gaussians—with diverse learning paradigms—including 2D-supervised learning, implicit neural representations, and 4D modeling—and connects them to real-world application scenarios. By constructing a structured knowledge graph of 3D vision, the study systematically relates dataset design, supervision mechanisms, and task requirements, clarifying the trade-offs between efficiency and fidelity and charting pathways for multimodal geometric grounding. This framework offers systematic guidance for reconstruction, generation, and dynamic scene modeling, advancing the field toward a unified and efficient paradigm.