核心概念
Jigsaw++ is a novel method that leverages the power of image-to-3D mapping and rectified flow to estimate complete 3D shape priors from partially assembled objects, thereby enhancing object reassembly tasks.
統計
Jigsaw++ reduces reconstruction errors on the Breaking Bad dataset, achieving a Chamfer Distance (CD) of 4.5e-2 compared to 10.5e-2 for the baseline Jigsaw method.
On PartNet, Jigsaw++ significantly improves precision and recall metrics for shape completion, exceeding the baseline DGL method by substantial margins across chair, table, and lamp categories.
When tested with 20% missing pieces on the Bottle category of the Breaking Bad dataset, Jigsaw++ maintains a low CD of 2.0e-2 and precision and recall of approximately 59.4%.
Augmenting the Jigsaw algorithm with Jigsaw++'s generated shape priors during global alignment reduces Jigsaw's error by 50%.