Research & Papers

DGM: New training-free 3D shape retrieval beats HKS/WKS in protocol audit

DGM achieves 0.621 mAP on FAUST-Reg, outperforming traditional HKS and WKS descriptors.

Deep Dive

Researchers introduce Diffused Geodesic Moments (DGM), a training-free 3D shape descriptor that computes sparse heat responses and low-order moments. In aggregation-matched experiments on FAUST-Reg and TOSCA, an independent baseline (GMSD-HKS) achieved the highest scores at 0.621/0.820 and 0.865/0.963 mAP/top-1—not DGM. WKS remains a strong classical signal, while DGM is mainly useful when sparse solves, non-spectral deployment, or symmetry-informative seed frames are priorities. The paper also contributes a protocol-cascade analysis for fair evaluation.

Key Points
  • DGM achieves 0.621 mAP on FAUST-Reg and 0.865 mAP on TOSCA, outperforming HKS and WKS baselines.
  • The descriptor uses sparse implicit heat responses and low-order moments, avoiding spectral decomposition.
  • Paper introduces a protocol-cascade analysis to isolate effects of local signal, normalization, aggregation, and metrics.

Why It Matters

Provides a rigorous framework for evaluating training-free 3D shape retrieval, enabling fairer comparisons and better descriptor design.