FDSML
Official models and data for FDSML: Formula-Driven Supervised Metric Learning with Parameter-Aware Triplet Loss (WACV 2027).
Hina Otake*, Shinei Arakawa*, Keitaro Tanaka*, Yoshihiro Fukuhara, Hirokatsu Kataoka, Shigeo Morishima (*equal contribution)
FDSML pre-trains vision models on synthetic fractal images only, without any natural images. Every image in FineFractalDB is paired with the parameters that generated it, and the Parameter-Aware Triplet Loss uses these parameters to shape the embedding space. With ResNet-50, FDSML raises Recall@1 over FDSL pre-training from 65.6% to 74.0% on Stanford Online Products and from 56.2% to 78.4% on Stanford Cars, while staying competitive on classification.
- Paper: https://arxiv.org/abs/2610.39423
- Project page: https://shineiarakawa.github.io/fdsml/
- Code: https://github.com/ShineiArakawa/fdsml
Dataset
- FineFractalDB: 1,000 IFS classes × 1,000 images with per-image rendering parameters.
Pre-trained models
FDSML models for each architecture and triplet selection strategy (all / each / same / diff, see the paper):
| Architecture | all |
each |
same |
diff |
|---|---|---|---|---|
| ResNet-50 | resnet50-all | resnet50-each | resnet50-same | resnet50-diff |
| ViT-Tiny (DeiT-Tiny/16, 224) | vittiny-all | vittiny-each | vittiny-same | vittiny-diff |
| ViT-Base (DeiT-Base/16, 224) | vitbase-all | vitbase-each | vitbase-same | vitbase-diff |
All checkpoints are plain PyTorch files and are loaded with the official code.
Citation
@inproceedings{Otake2027fdsml,
title = {FDSML: Formula-Driven Supervised Metric Learning with Parameter-Aware Triplet Loss},
author = {Hina Otake and Shinei Arakawa and Keitaro Tanaka and Yoshihiro Fukuhara and Hirokatsu Kataoka and Shigeo Morishima},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
year = {2027},
}