The paper addresses the challenge of generating object-centric layout designs under spatial constraints, particularly in the context of floorplan design. The key contributions are:
Representation: Each room is represented using four latent variables (xmin, xmax, ymin, ymax) to capture fine-grained interactions between rooms, unlike the typical single-node representation.
Factor Graph Model: A factor graph is constructed to effectively model the dependencies between the room variables, including higher-order constraints. This allows the incorporation of domain knowledge about spatial relationships.
Factor Graph Neural Network (FP-FGNN): The factor graph is converted into a neural network that performs message passing to learn the optimal room coordinates, which are then used to generate the final floorplan layout.
The proposed approach outperforms existing methods on both box-level and pixel-level metrics, demonstrating its ability to generate layouts that closely match the ground truth. It also shows strong performance in practical scenarios, such as iterative design with partial user inputs and diverse generation from the same boundary.
翻譯成其他語言
從原文內容
arxiv.org
深入探究