FDSML: Formula-Driven Supervised Metric Learning with Parameter-Aware Triplet Loss
On Stanford Cars, FDSML raised retrieval Recall@1 from 56.2% to 78.4% over FDSL pre-training using ResNet-50.
The method learns image similarities from the parameters that generate synthetic fractals, rather than relying on discrete class labels. It combines a refined fractal dataset with a triplet loss that selects positive and negative examples by their distances in generation-parameter space. Stanford Online Products Recall@1 also rose from 65.6% to 74.0%. The authors report competitive downstream classification and improved robustness under several corruption settings.
HF Daily Papers' note
The method learns image similarities from the parameters that generate synthetic fractals, rather than relying on discrete class labels. It combines a refined fractal dataset with a triplet loss that selects positive and negative examples by their distances in generation-parameter space. Stanford Online Products Recall@1 also rose from 65.6% to 74.0%. The authors report competitive downstream classification and improved robustness under several corruption settings.
HF Daily Papers' note
score 4