Technical gazette, Vol. 33 No. 2, 2026.
Original scientific paper
https://doi.org/10.17559/TV-20240908001975
Enhanced Structure-from-Motion 3D Reconstruction through Deep Learning Feature Fusion and Optimization
Dong Li
; China Railway Smart City Research and Development Center, China Railway Liuyuan Group Co., Ltd. No. 36 Zhonghuan West Road, Tianjin Airport Economic Area, Tianjin, China
Gongyun Fu
; China Railway Smart City Research and Development Center, China Railway Liuyuan Group Co., Ltd. No. 36 Zhonghuan West Road, Tianjin Airport Economic Area, Tianjin, China
Shunjie Yang
; School of Software Engineering, Beijing Jiaotong University, No. 3 Shangyuancun, Haidian District, Beijing, China
*
* Corresponding author.
Abstract
This paper presents Mixum, a novel 3D reconstruction framework for Structure-from-Motion (SfM), which combines traditional feature extraction and matching techniques with deep learning-based optimization. The Mixum framework enhances the accuracy of feature matching and eliminates redundant feature points. Additionally, the integration with PixSfM, a deep-learning accuracy optimization algorithm, further reduces reprojection error and enhances multi-view consistency. Experiments on multiple public datasets reveal that Mixum significantly improves 3D reconstruction density and reduces reprojection error by up to 23%, demonstrating its applicability for complex scenes in applications like cultural heritage preservation, virtual reality, and autonomous navigation.
Keywords
3D reconstruction; attention mechanism; convolutional neural network; deep learning; feature extraction and matching
Hrčak ID:
344967
URI
Publication date:
28.2.2026.
Visits: 488 *