Topology-Inspired Fine-Tuning for Wind-Field Super-Resolution
The combined fine-tuning objective cut three topology error measures by 25–31%, while delivering about half the pixel-fidelity gain of standard fine-tuning.
On a fixed benchmark of 168 wind fields, the method improved all three persistence-diagram distances on 150 fields relative to the pretrained CNN. It combines gradient supervision with constraints on values and persistence at ground-truth feature-pair endpoints. Mean merge-tree distance also improved, but only on 91 of those 150 fields—a limit showing that the two topology descriptors capture complementary structure. Standard fine-tuning achieved a larger PSNR gain but worsened all three persistence-diagram distances.
ArXiv · AI/CL/LG's note
On a fixed benchmark of 168 wind fields, the method improved all three persistence-diagram distances on 150 fields relative to the pretrained CNN. It combines gradient supervision with constraints on values and persistence at ground-truth feature-pair endpoints. Mean merge-tree distance also improved, but only on 91 of those 150 fields—a limit showing that the two topology descriptors capture complementary structure. Standard fine-tuning achieved a larger PSNR gain but worsened all three persistence-diagram distances.
ArXiv · AI/CL/LG's note
score 4