Abstract
We present Raster2Hull, a deep learning model for reconstructing 3D ship hulls directly from 2D raster general arrangement and lines plans. Although modern shipbuilding is heavily digitized, many aging vessels nearing the end of their service lives lack 3D models, hindering downstream applications such as recycling planning and digital twin construction. To recover 3D hull geometry, our architecture first processes the raster drawings through a shared convolutional encoder and segmentation head. Instead of relying on standard multi-view pooling, our method employs a dynamic tokenizer that converts these localized features into sparse semantic tokens, which then drive a transformer-based refiner to deform a template 3D hull into the final point cloud. Furthermore, we introduce a domain-specific, multi-objective loss formulation incorporating a spatially weighted Chamfer distance to capture complex extremities (e.g., the bow and stern), coupled with normal consistency. Addressing the scarcity of public benchmarks in naval architecture, we release ShipPlan, a synthetic dataset comprising 5,000 paired 2D raster drawings and 3D point clouds across diverse hull topologies. Evaluations against representative baselines, supported by internal ablations of our loss formulation, indicate that the framework successfully captures global hull topology from the provided general arrangement and lines plans.