CUT3R-Marine: Persistent-State 3D Reconstruction with Learnable Filtering for Offshore Environments

Extend CUT3R’s persistent-state 3D perception (Wang et al., 2025) by embedding a learnable filtering module that generalizes Ran et al.’s KNN_PCF + hyperbolic-weight hybrid filter (2024) as differentiable layers. The system ingests streaming LiDAR/photogrammetry in windy, wavy conditions, learns to filter dynamic noise online, and reconstructs a coherent scene in a shared coordinate frame. ALSO (Boulch et al., 2022) pretraining on occupancy estimation provides a strong surface-inductive prior. Current offshore filtering is algorithmic and separate from reconstruction. Here, filtering is learned, end-to-end, inside a persistent-state model that updates with each new observation. The model also learns a noise generative prior specialized to marine conditions (e.g., spectral patterns of motion-induced outliers). This combines three strands rarely integrated: (i) stateful reconstruction with virtual probing (CUT3R), (ii) hand-engineered marine filters (Ran et al.) turned into trainable modules, and (iii) self-supervised surface pretraining (ALSO). DUSt3R (Wang et al., 2023) can serve as an initialization for pose-free pairwise reconstruction, with CUT3R consolidating globally. Offshore assets are huge, dynamic, and safety-critical. An end-to-end system tuned to real noise statistics should preserve fine structural details while removing transient clutter, lowering post-processing burden and error. Impact includes faster, more accurate digital twins for inspection and prefabrication, enabling frequent, reliable updates in challenging field conditions.

References:

1. Combined Filtering Method for Offshore Oil and Gas Platform Point Cloud Data Based on KNN_PCF and Hy_WHF and Its Application in 3D Reconstruction. Chunqing Ran, Xiaobo Zhang, Haobin Yu, Zhengyang Wang, Shengli Wang, Jichao Yang (2024). Italian National Conference on Sensors.
2. ALSO: Automotive Lidar Self-Supervision by Occupancy Estimation. Alexandre Boulch, Corentin Sautier, Bjoern Michele, Gilles Puy, Renaud Marlet (2022). Computer Vision and Pattern Recognition.
3. Continuous 3D Perception Model with Persistent State. Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, Angjoo Kanazawa (2025). Computer Vision and Pattern Recognition.
4. DUSt3R: Geometric 3D Vision Made Easy. Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, Jérôme Revaud (2023). Computer Vision and Pattern Recognition.