Enabling Multi-scale Features and Double-Attention for Single-image 3D Urban Reconstruction

Miao, Jun; Shen, Jian; Jiang, Ruixiang; Chu, Jun; Guo, Luanzheng

doi:10.1109/cac53003.2021.9728424

Cited by 1 publication

(1 citation statement)

References 21 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…While Pix2Point successfully predicts point clouds for 3D scenes from single outdoor images, the resulting outdoor 3D scene reconstructions often suffer from limited visual fidelity. In another endeavor, Jun Miao et al proposed EMERALD [8] , a 3D reconstruction network founded on single images. EMERALD integrates multiscale contextual features, incorporating (DSC ASPP), into a fully convolutional network (FCN).…”

Section: Ground-based Capture Image Reconstructionmentioning

confidence: 99%

Global Soft Pooling Adaptive Attention Network for Single Aerial Image-Based 3D Outdoor Scene Reconstruction

Wang,

Li,

Yan

2023

Preprint

View full text Add to dashboard Cite

In the realm of outdoor scene 3D reconstruction from a solitary RGB aerial image, a persistent challenge pertains to the issue of inadequate reconstruction accuracy. In this article, we introduce a novel approach, namely, the Global Soft Pooling Adaptive Attention Network, designed to facilitate high-precision 3D scene reconstruction in outdoor environments. This network primarily comprises two distinct attention network modules: the Global Soft Pooling Dynamic Convolutional Attention and the Three-Head Adaptive Graph Attention. The Global Soft Pooling Attention Network employs a stem_conv and four MBdyconv components for the extraction of multi-scale image features. It harnesses global soft pooling techniques, integrating both channels fusion and spatial attention mechanisms, thereby enabling the precise acquisition of composite features. These features are subsequently assigned to initialize mesh vertices. The Three-Head Adaptive Graph Attention Network leverages three sets of adaptive 1D convolutions to derive 3D vertex features, taking into account the weights of neighboring vertices. A linear layer is employed to compute the vertex coordinate offsets (denoted as "ΔV"), which are then utilized to refine the initial mesh vertices, ultimately achieving an enhanced mesh model. When evaluated on the publicly available SensatUrban dataset, our proposed method demonstrates impressive performance metrics, achieving a reconstruction performance index of 0.96 for l2 and 1.68 for l3. Experimental results unequivocally demonstrate that our approach outperforms existing deep learning methodologies, establishing itself as the state-of-the-art solution for outdoor 3D scene reconstruction.

show abstract

Section: Ground-based Capture Image Reconstructionmentioning

confidence: 99%