Epipolar-Constrained Cross-Attention for Triangulation-Guided Depth Estimation From Multi-View Images
IEEE Access · 2026 · 14 · 108866-108878
DOI: 10.1109/access.2026.3713483 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Okhwan Bae und 1 weiteren Autor:innen. Die Arbeit nennt 13 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Institute of Information and Communications Technology Planning and Evaluation (IITP)-Innovative Human Resource Development for Local Intellectualization progr…, Korea Government [Ministry of Science and ICT (MSIT)].
Abstract (Originalsprache)
Back Projection, a fundamental process in 3D reconstruction, requires depth information to restore 2D pixels to 3D coordinates. However, since depth information is lost during the projection of 3D space onto a 2D plane, accurate reconstruction is impossible using geometric formulas alone. To address this limitation, this paper proposes a novel deep learning approach that infers lost depth information by analyzing the relationships between multi-view-images using artificial intelligence, thereby reconstructing high-resolution depth maps. The proposed method embeds reference and target images into 3D voxel feature grids using an encoder-decoder architecture based on a Convolutional Neural Network, and restores depth information via a cross-attention mechanism. In particular, this study advances the attention process by applying epipolar geometry to the front plane of the 3D voxel grid. Unlike conventional methods that search all pixels globally, our model generates an epipolar line corresponding to the query of the reference image on the front face of the 3D voxel grid and limits the search scope to features sampled along this line. Consequently, the query performs attention only on the keys and values of the target image that exist on the geometrically valid epipolar line, thereby reducing unnecessary computations and ensuring the reliability of depth estimation. Furthermore, to refine the geometric information estimated at low resolution into high-resolution, we introduce a Smart Upscaling technique. The proposed model first computes a reliable Geometric Depth at a reduced resolution using triangulation based on epipolar geometry and attention maps. Subsequently, a separate upscale decoder receives these features and progressively restores the resolution. This coarse-to-fine approach effectively suppresses noise even in texture-less or complex regions, enabling the generation of sharp, high-resolution depth maps. Experimental results on the NeRF Synthetic dataset demonstrate that the proposed model significantly outperforms contemporary Multi-View Stereo baseline models. Specifically, the proposed method achieved an L1 Error of 0.0298 and an Accuracy of 86.21%, consistently surpassing both the Cascade Multi-View Stereo Network (L1 Error: 0.0361, Accuracy: 81.87%) and the high-performance Transformer-based Multi-View Stereo Network (L1 Error: 0.0339, Accuracy: 84.72%). These findings confirm that our approach successfully recovers high-fidelity depth maps with superior geometric precision, closely aligning with the ground truth even in structurally complex scenes. Through these achievements, this study presents a novel deep learning methodology that leverages Epipolar Geometry and Cross-Attention to achieve high-precision depth estimation, thereby establishing a robust framework that effectively synthesizes geometric constraints with data-driven inference.
Förderung
- Institute of Information and Communications Technology Planning and Evaluation (IITP)-Innovative Human Resource Development for Local Intellectualization progr…
- Korea Government [Ministry of Science and ICT (MSIT)]
Zitieren
APA
Bae, O., & Hong, C. (2026). Epipolar-Constrained Cross-Attention for Triangulation-Guided Depth Estimation From Multi-View Images. IEEE Access, 14, 108866-108878. https://doi.org/10.1109/access.2026.3713483
Vancouver
Bae O, Hong C. Epipolar-Constrained Cross-Attention for Triangulation-Guided Depth Estimation From Multi-View Images. IEEE Access. 2026;14:108866-108878. doi:10.1109/access.2026.3713483
BibTeX
@article{Bae2026,
author = {Bae, Okhwan and Hong, Chung-Pyo},
title = {Epipolar-Constrained Cross-Attention for Triangulation-Guided Depth Estimation From Multi-View Images},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {108866--108878},
doi = {10.1109/access.2026.3713483}
}
EUVIMED