Implicit 3D Reconstruction of Fine Details from Multi-View Images using Wavelet-based Geometric Prior
Abstract: High-fidelity 3D reconstruction from images remains a fundamental challenge in computer vision. Implicit Signed Distance Field (SDF) models leverage photometric loss for isosurface reconstruction, while recent approaches, such as planar constrained Gaussian splatting, integrate 3D-2D geometry priors to improve structural accuracy. However, existing methods struggle to capture fine-grained geometric details due to due the loss of high-frequency geometric details during feature learning, which results in limited multi-scale representation. To address this, we introduce a novel wavelet-conditioned implicit SDF model that enhances geometric precision by leveraging a pretrained wavelet autoencoder optimized with sharp depth maps. This autoencoder extracts multi-scale wavelet transformed features, which are fused with implicit 3D triplane features via triplane projection, producing a more structured and detail-preserving distance field. Our method can serve as a plug-and-play module, seamlessly integrating with any implicit SDF representations.
Extensive evaluations on DTU, Tanks and Temples, and a cultural heritage dataset demonstrate that our model consistently outperforms state-of-the-art implicit and explicit 3D reconstruction methods, achieving more complete surfaces with fine-detail preservation across diverse scene scales, from small objects to large architectural buildings.
Submission Length: Long submission (more than 12 pages of main content)
Assigned Action Editor: ~Jaesik_Park3
Submission Number: 4391
Loading