Automated culvert inspection systems can help increase the safety and efficiency of flood management operations. As a key step to this system, we present an efficient RGB-based 3D reconstruction pipeline for culvert-like structures in visually repetitive environments. Our approach first selects informative frame pairs to maximize viewpoint diversity while ensuring valid correspondence matching using a plug-and-play module, followed by a reconstruction model that simultaneously estimates RGB appearance, geometry, and semantics in real-time. Experiments demonstrate that our method effectively generates accurate 3D reconstructions and depth maps, enhancing culvert inspection efficiency with minimal human intervention.
Figures & tables
Figure 1: Camera pose alignment. For each scene, ground truth poses were scaled and aligned to match the coordinate system predicted by the reconstruction model.
Figure 2: Application of pairwise selection algorithm to LSM . For each scene, two informative frames are selected as input to the reconstruction model, which simultaneously predicts RGB appearance, depth map, and feature field for a smooth interpolated trajectory that includes novel views.
Figure 3: Application of pairwise selection algorithm to Baseline models . The pairwise selection algorithm can help to increase the potential for successful scene reconstruction of novel views for various models in general.
Image Size
Peak GPU Memory (GB)
Inference Time (sec)
Total Execution Time (sec)
512×288
14.04
0.525
59.66
256×144
8.95
0.487
48.02
Table 1: Average inference efficiency of model. The additional frame selection module shows minimal effect on the efficiency of 3D reconstruction.
Selection Strategy
PSNR ↑
SSIM ↑
LPIPS ↓
Random Selection
5.8412
0.0912
0.77434
First and Last
6.3638
0.1156
0.73582
Q1 and Q3
5.4122
0.0504
0.81794
Ours
14.124
0.1924
0.45808
Table 2: Comparison of different frame pair selection strategies for input used to train LSM.