FDSML uses formula parameters for metric learning without real images
Original title: FDSML: Formula-Driven Supervised Metric Learning with Parameter-Aware Triplet Loss
On ResNet-50, Recall@1 rises from 65.6% to 74.0% on Stanford Online Products and from 56.2% to 78.4% on Stanford Cars versus FDSL.
Formula-Driven Supervised Metric Learning (FDSML) replaces discrete class labels in synthetic vision pre-training with relational supervision from image-generation parameters. It pairs FineFractalDB, a fractal dataset with higher intra-class variation, with a Parameter-Aware Triplet loss that ranks positives and negatives by distance in Iterated Function System parameter space. The paper reports better transfer to image retrieval, competitive classification, and stronger robustness under several corruption settings.
- FineFractalDB is described as a refined fractal set with more stable structures.
- PA-Triplet loss trains embeddings to match parameter-derived similarity orderings.
- Retrieval gains are reported over FDSL pre-training on ResNet-50.
- Classification stays competitive while corruption robustness improves.
- Products
- FineFractalDBResNet-50
From the original article
1 min read at huggingface.co
Formula-Driven Supervised Learning (FDSL) enables vision pre-training without real images, but existing methods cast pre-training as classification on discretized labels, discarding continuity in the generative process. We propose Formula-Driven Supervised Metric Learning (FDSML), which derives relational supervision from image-generation parameters beyond discrete class labels.
By Hina Otake, Shinei Arakawa, Keitaro Tanaka · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.