BROWSE
Detail
Conference Paper
Refining Image-to-3D Foundation Models via Geometric Supervision for Industrial Plant Reconstruction
Cited - time in 
Share 
Share
- Authors
- SangEun Lee, Wonseok Chae, Hoyoung Yoo, Geunyong Kim, NackWoo Kim, Hyeonjin Kim
- Issue Date
- 2026-06
- Citation
- Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) 2026, pp.9194-9202
- Publisher
- Computer Vision Foundation
- Language
- English
- Type
- Conference Paper
- Abstract
- Recent image-to-3D foundation models generate visually plausible objects from single images, yet they often fail to satisfy the geometric strictness required in industrial environments. In plant facility reconstruction, where structures are thin, layouts are strictly orthogonal, existing methods frequently produce duplicated structures and unstable thin elements. These limitations reduce their applicability in real-world scenarios such as digital twins, automated maintenance, and facility monitoring. In this paper, we present a two-stage geometry-constrained adaptation framework for a pretrained image-to-3D foundation model. The first stage supervises part-level normal vectors to stabilize local geometry and enhance planar and orthogonal alignment. The second stage leverages Signed Distance Field guidance on whole-object data, ensuring global consistency, eliminating spurious surfaces, and promoting watertight topology. This strategy strengthens geometric faithfulness without modifying the underlying generative architecture. Experiments on curated industrial benchmarks adapted for plant facility reconstruction demonstrate consistent improvements in structural, appearance, and semantic fidelity, showing that explicit geometric constraints enable generative 3D foundation models to meet the demands of high-precision industrial reconstruction.
- KSP Keywords
- Automated maintenance, First stage, Global consistency, High-precision, Industrial plant, Object Data, Real-world, Two-Stage, adaptation framework, facility monitoring, geometric constraints