研究目的
To present a unified view of three fidelities between measures (Energy Distance, weighted Hausdorff distance, Wasserstein distance) that alleviate the problem of semi-local gradients in shape registration pipelines, and to implement them through efficient GPU routines.
研究成果
The paper introduces three positive divergences (Energy Distance, ε-Sinkhorn cost, ε-Hausdorff divergence) for shape analysis, providing theoretical guarantees and efficient GPU routines. These tools are designed to improve the robustness of shape registration pipelines to large deformations.
研究不足
The paper does not explicitly mention limitations, but the reliance on GPU routines may limit accessibility for users without compatible hardware.
1：Experimental Design and Method Selection:
The paper leverages theoretical advances on Sinkhorn entropies and divergences to unify three fidelities between measures. It implements these through efficient GPU routines.
2：Sample Selection and Data Sources:
The focus is on the registration of normalized density maps, represented as measures on a feature space.
3：List of Experimental Equipment and Materials:
The experiments utilize GPU routines with Matlab, numpy, and pytorch bindings, specifically the KeOps library for efficient computation.
4：Experimental Procedures and Operational Workflow:
The methodology involves computing gradients of the unified fidelities to drive registration algorithms, with a focus on handling large deformations robustly.
5：Data Analysis Methods:
The analysis involves comparing the performance and robustness of the proposed fidelities in shape registration tasks.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
