The article introduces PARTSTAD, a method designed for the task adaptation of 2D-to-3D segmentation lifting. It optimizes a 2D bounding box prediction model for 3D segmentation, introducing weights for 2D bounding boxes for adaptive merging and learning the weights using a small additional neural network. The method also incorporates SAM, a foreground segmentation model, to improve the boundaries of 2D segments and consequently those of 3D segmentation. The experiments on the PartNet-Mobility dataset show significant improvements in semantic and instance segmentation compared to the state-of-the-art few-shot 3D segmentation model.
Publication date: 11 Jan 2024
Project Page: https://arxiv.org/abs/2401.05906v1
Paper: https://arxiv.org/pdf/2401.05906
