The authors present a novel approach for partial-to-partial 3D shape matching that exploits geometric consistency as a strong constraint. Their key contributions are:
A geometrically consistent partial-to-partial shape matching formalism realized through an integer non-linear program. This fuses state-of-the-art deep features with a non-linear integer programming approach.
A pruned search algorithm to efficiently solve the integer program by iteratively solving a sequence of integer linear programs.
A new inter-class partial-to-partial dataset based on the SMAL dataset to expand the scope of partial-to-partial shape matching evaluation.
The authors show that their method outperforms current state-of-the-art supervised deep learning and combinatorial optimization algorithms in both intra-class and inter-class partial-to-partial shape matching settings, as evaluated by intersection over union (IoU) and geodesic error metrics.
The key technical innovation is the incorporation of geometric consistency constraints into the partial-to-partial matching problem, which was not previously explored. This allows the method to produce smooth, geometrically coherent matchings between partial shapes, going beyond existing approaches that either require full shape annotations or cannot handle partial-to-partial matching.
The authors also introduce a new inter-class partial-to-partial dataset to broaden the evaluation of this challenging problem. Overall, the work represents a significant advancement in addressing real-world partial-to-partial shape matching scenarios.
Til et andet sprog
fra kildeindhold
arxiv.org
Vigtigste indsigter udtrukket fra
by Viktoria Ehm... kl. arxiv.org 04-19-2024
https://arxiv.org/pdf/2404.12209.pdfDybere Forespørgsler