Research map / Label-efficient and robust learning / Self-supervised and contrastive
Discovering Visual Patterns in Art Collections With Spatially-Consistent Feature Learning
Xi Shen, Alexei A. Efros, Mathieu Aubry
Check 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 |
|---|---|---|---|
| Fine-tune ImageNet features using spatially consistent self-supervised matches on Brueghel | ResNet-18 features · Brueghel | mAP (cosine similarity): 58.0 -> 75.3 | Improved |
| Apply the discovery score instead of cosine similarity | Ours (trained on Brueghel) · Brueghel | mAP: 75.3 -> 76.4 | Improved |
| Train the feature on Brueghel rather than LTLL | ResNet-18 features · Brueghel | mAP (cosine similarity): 65.2 -> 75.3 | Improved |
| Train the feature on LTLL rather than ImageNet-only features | ResNet-18 features · Brueghel | mAP (cosine similarity): 58.0 -> 65.2 | Improved |
| Fine-tune features on LTLL and use discovery | ResNet-18 + discovery · LTLL | classification accuracy (%): 80.9 -> 88.5 | Improved |
| Use the discovery score instead of image-level max-pooled features | ResNet-18 · LTLL | classification accuracy (%): 59.8 -> 80.9 | Improved |
| Fine-tune features on Oxford5K | ResNet-18 + discovery · Oxford5K | retrieval mAP (%): 85.0 -> 85.7 | Improved |
| Train features on LTLL rather than Oxford5K | ResNet-18 + discovery · Oxford5K | retrieval mAP (%): 85.7 -> 83.6 | Hurt |
| Use Context Prediction training instead of spatial-consistency feature learning | ResNet-18 features · Brueghel | mAP (cosine similarity): 58.0 -> 58.8 | No clear effect |
| Use a positive region farther from the proposal region (P12 or P14) instead of a nearby positive region (ablation) | Spatially-consistent feature learning · Brueghel | one-shot matching mAP | Improved |
| Use positive regions close to the proposal region (ablation) | Spatially-consistent feature learning · Brueghel | one-shot matching mAP | Hurt |
Closest papers by meaning
- The Met Dataset: Instance-level Recognition for Artworks (NeurIPS 2021 · similarity 0.86)
- Sketching With Style: Visual Search With Sketches and Aesthetic Context (ICCV 2017 · similarity 0.84)
- A Deep Learning Approach to Clustering Visual Arts (IJCV 2022 · similarity 0.84)
- ALADIN: All Layer Adaptive Instance Normalization for Fine-Grained Style Similarity (ICCV 2021 · similarity 0.83)
- Progressive Neighbor Consistency Mining for Correspondence Pruning (CVPR 2023 · similarity 0.83)
- Disentangling Structure and Aesthetics for Style-Aware Image Completion (CVPR 2018 · similarity 0.83)
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.
