Research on Monocular Image 3D Reconstruction Technology Based on End-to-End Deep Neural Networks and Neural Rendering

Abstract

The task of monocular three-dimensional reconstruction represents a critical yet fundamentally ill-posed inverse problem in the field of computer vision, constrained primarily by the inherent loss of depth information during the two-dimensional perspective projection process. Existing optimization-based neural rendering methodologies frequently struggle with reconstructing high-fidelity geometric structures and handling severe spatial occlusions from a single viewpoint, while simultaneously suffering from prohibitive computational overhead during inference that strictly limits their applicability in real-time deployment scenarios. To address these limitations, this work develops an end-to-end feed-forward reconstruction framework that combines pixel-aligned feature extraction, linearcomplexity state-space modeling, and explicit Gaussian rendering. Rather than treating these established components as independent contributions, the proposed method introduces a global-context-before-lifting strategy: long-range dependencies are first modeled over image-space features, after which the contextualized features are restored to their spatial locations and directly decoded into ray-conditioned 3D Gaussian primitives. This design enables each primitive to exploit both local appearance and global structural cues while preserving pixel-toray correspondence, allowing single-image reconstruction without test-time gradient descent or per-scene optimization. Extensive experimental evaluations and rigorous ablation studies conducted on the standard ShapeNet, DTU, and ScanNet benchmark datasets comprehensively demonstrate that the proposed feed-forward method achieves competitive quantitative performance across Peak Signal-to-Noise Ratio, Structural Similarity Index Measure, and Chamfer Distance metrics. Furthermore, by replacing implicit volumetric numerical integration with analytical alpha blending, the network simultaneously enables real-time rendering capabilities exceeding sixty frames per second during inference. Overall, the proposed framework provides an efficient solution for synthesizing geometrically consistent and photorealistic novel views from monocular inputs.