Anomaly-Amplifying Reconstruction: Preserving the Outliers Instead of Smoothing Them Away

Standard reconstruction-based 3D anomaly detectors tend to “repair” anomalies, making them harder to detect. Building on Liang et al.’s DUS-Net (2025), which preserves group-center structures in high-precision point clouds, this project proposes a dual-branch reconstruction model: a normalizer branch that reconstructs the canonical, anomaly-free surface (as in DUS-Net), and a residual-preservation branch that is explicitly optimized to retain and amplify local deviations from the canonical surface. The two branches are trained with complementary losses, where the normalizer is penalized for fitting anomalies and the residual branch is rewarded for reconstructing them with correct geometry and depth. Current methods (e.g., DUS-Net) aim for high-fidelity reconstructions and use reconstruction error as an anomaly cue, but they still optimize away subtle defects. Here, anomalies are promoted to first-class citizens: the model learns to predict a factored representation (normal surface + anomaly residual) with multi-scale geometric priors that prevent residual over-smoothing. It takes the “group center preserving” idea of DUS-Net and inverts the usual objective by designing a residual-preserving channel with anomaly-consistency constraints. It also leverages rail-surface work by Hu et al. (2024) to evaluate local depth accuracy of anomalies after multi-view reconstruction, but targets better retention and quantification of shallow defects (sub-mm) that PatchmatchNet-based pipelines may partially smooth away. DUSt3R’s pointmap regression (Wang et al., 2023) can be used as a backbone to avoid camera calibration overhead in industrial settings. Safety-critical domains (rails, aerospace, additive manufacturing) need accurate 3D geometry of tiny defects. Explicit anomaly-preserving reconstruction should reduce false negatives and improve depth estimation for maintenance decisions (as shown critical in Hu et al., 2024). Impact includes better inspection accuracy with more faithful defect geometry; fewer misses on early-stage damage; principled separation of “expected” vs. “unexpected” geometry that can transfer across materials and sensors.

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. DUSt3R: Geometric 3D Vision Made Easy. Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, Jérôme Revaud (2023). Computer Vision and Pattern Recognition.