Research map / Label-efficient and robust learning / Recognition and long-tail
ALADIN: All Layer Adaptive Instance Normalization for Fine-Grained Style Similarity
Dan Ruta, Saeid Motiian, Baldo Faieta, Zhe Lin, Hailin Jin, Alex Filipkowski, Andrew Gilbert, John Collomosse
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add contrastive loss instead of triplet loss (ablation) | ALADIN-L · BAM-FG | IR Top-1: 26.70 -> 56.89 | Improved |
| add contrastive loss instead of listwise loss (ablation) | ALADIN-L · BAM-FG | IR Top-1: 37.90 -> 56.89 | Improved |
| add weak group supervision instead of reconstruction only (ablation) | ALADIN-L · BAM-FG | IR Top-1: 11.48 -> 56.89 | Improved |
| use ALADIN architecture instead of MUNIT style code under contrastive training | contrastive training · BAM-FG | IR Top-1: 6.43 -> 56.89 | Improved |
| use larger ALADIN-L backbone instead of ALADIN-S (ablation) | ALADIN contrastive · BAM-FG | IR Top-1: 36.42 -> 56.89 | Improved |
| add projection MLP to listwise loss (ablation) | ALADIN-L listwise · BAM-FG | IR Top-1: 25.14 -> 37.90 | Improved |
| add projection MLP to triplet loss (ablation) | ALADIN-L triplet · BAM-FG | IR Top-1: 26.70 -> 22.12 | Hurt |
| add hard negative mining to triplet loss (ablation) | ALADIN-L triplet · BAM-FG | IR Top-1: 26.70 -> 23.75 | Hurt |
| add group supervision instead of SimCLR-style augmentation (ablation) | ALADIN-L contrastive · BAM-FG | IR Top-1: 30.47 -> 56.89 | Improved |
| add group supervision instead of augmentation (ablation) | ResNet50 contrastive · BAM-FG | IR Top-1: 28.51 -> 40.14 | Improved |
| increase training data from 250K to 1M images (ablation) | ALADIN-L contrastive · BAM-FG | IR Top-1: 40.08 -> 56.89 | Improved |
| fine-tune on cleaned BAM-FG-C5 data (ablation) | ALADIN-L · BAM-FG | IR Top-1: 56.89 -> 58.98 | Improved |
Closest papers by meaning
- Sketching With Style: Visual Search With Sketches and Aesthetic Context (ICCV 2017 · similarity 0.90)
- StyleBabel: Artistic Style Tagging and Captioning (ECCV 2022 · similarity 0.89)
- Investigating Style Similarity in Diffusion Models (ECCV 2024 · similarity 0.88)
- A Learned Representation For Artistic Style (ICLR 2017 · similarity 0.87)
- FiVA: Fine-grained Visual Attribute Dataset for Text-to-Image Diffusion Models (NeurIPS 2024 · similarity 0.86)
- Stylebreeder: Exploring and Democratizing Artistic Styles through Text-to-Image Models (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
