DVS-Defect: Differentiable 2D-to-3D Supervision for Depth of Industrial Anomalies

Inspired by Luo et al.’s differentiable voxelization/slicing (DVS, 2024) for cardiac mesh reconstruction, this work supervises 3D defect geometry using only 2D inspection measurements—e.g., single-shot fringe patterns (Nguyen & Wang, 2024) or thermal maps—by backpropagating slice-level losses to a 3D mesh/implicit surface. The network learns a 3D representation whose differentiable projections reconstruct measured 2D signals across multiple modalities and viewpoints. Industrial 3D defect pipelines (e.g., Hu et al., 2024) depend on dense point clouds or Patchmatch-style MVS for depth. Here, supervision is inverted: fitting 3D geometry to match 2D evidence via an explicit differentiable rendering/slicing bridge, removing the need for highly accurate point clouds in early training or domain shifts across sensors. This combines DVS (Luo et al., 2024) with single-shot fringe learning (Nguyen & Wang, 2024), where the predicted 3D surface must reproduce the fringe responses at multiple frequencies. A reconstruction-based anomaly prior (as in Liang et al., 2025) can be incorporated to enforce plausible normal surfaces while letting residuals account for defects. This creates a shared training framework for 3D reconstruction from diverse 2D industrial signals, enabling cheaper, faster acquisition. Cross-domain supervision reduces reliance on dense 3D ground truth and can generalize to new materials. Impact includes a practical path to sub-mm defect depth estimation using commodity 2D sensors, improving deployment robustness in factory and rail environments while cutting measurement time and cost.

References:

1. 3D reconstruction and depth estimation method for local anomalies of rail surface based on multi-view stereo matching. Pengyu Hu, Qianwen Zhong, Shubin Zheng, Xieqi Chen, Lele Peng (2024). Measurement science and technology.
2. Taming Anomalies with Down-Up Sampling Networks: Group Center Preserving Reconstruction for 3D Anomaly Detection. Hanzhe Liang, Jie Zhang, Tao Dai, Linlin Shen, Jinbao Wang, Can Gao (2025). arXiv.org.
3. Explicit Differentiable Slicing and Global Deformation for Cardiac Mesh Reconstruction. Yihao Luo, Dario Sesia, Fanwen Wang, Yinzhe Wu, Wenhao Ding, Jiahao Huang, Fadong Shi, A. Shah, Amit Kaural, J. Mayet, Guang Yang, C. Yap (2024). arXiv.org.
4. Single-Shot 3D Reconstruction via Nonlinear Fringe Transformation: Supervised and Unsupervised Learning Approaches. Andrew-Hieu Nguyen, Zhaoyang Wang (2024). Italian National Conference on Sensors.