Research map / Label-efficient and robust learning / Self-supervised and contrastive
Benchmarking Representation Learning for Natural World Image Collections
Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, Oisin Mac Aodha
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 |
|---|---|---|---|
| pretrain with supervised labels on iNat2021 instead of ImageNet | ResNet50 feature extractor · NeWT (164-task mean) | top-1 accuracy: 0.744 -> 0.806 | Improved |
| use iNat2021 full training data instead of iNat2021 mini | supervised ResNet50 feature extractor · NeWT (164-task mean) | top-1 accuracy: 0.793 -> 0.806 | Improved |
| use iNat2021 full training data instead of iNat2021 mini | SimCLR feature extractor · NeWT (164-task mean) | top-1 accuracy: 0.678 -> 0.689 | Improved |
| pretrain with self-supervision on iNat2021 rather than ImageNet | self-supervised feature extractors · CUB | top-1 accuracy: 0.631 -> 0.365 | Hurt |
| use ImageNet SwAV rather than supervised ImageNet features | ResNet50 feature extractor · StanfordCars | top-1 accuracy: 0.479 -> 0.595 | Improved |
| use self-supervised rather than supervised ImageNet features | feature extractor · NeWT Gestalt | top-1 accuracy: 0.859 -> 0.908 | Improved |
| use self-supervised rather than supervised ImageNet features | feature extractor · NeWT Appearance - Species | top-1 accuracy: 0.725 -> 0.647 | Hurt |
| use iNat2021 mini SimCLR x4 instead of iNat2021 mini SimCLR | ResNet50 feature extractor · Flowers102 | top-1 accuracy: 0.814 -> 0.911 | Improved |
| initialize supervised training randomly instead of from ImageNet (ablation) | ResNet50 trained on iNat2021 full · iNat2021 test | top-1 accuracy: 0.760 -> 0.746 | Hurt |
| initialize supervised training randomly instead of from ImageNet (ablation) | ResNet50 trained on iNat2021 mini · iNat2021 test | top-1 accuracy: 0.654 -> 0.616 | Hurt |
| add a geographical prior trained on iNat2021 to the iNat2021 vision model | supervised ResNet50 · iNat2021 test | top-1 accuracy: 0.760 -> 0.800 | Improved |
| use a geographical prior trained on full iNat2021 rather than mini | vision iNat2021 mini + geographical prior · iNat2021 test | top-1 accuracy: 0.722 -> 0.739 | Improved |
Closest papers by meaning
- Focus on the Positives: Self-Supervised Learning for Biodiversity Monitoring (ICCV 2021 · similarity 0.91)
- Dual-Margin Embedding for Fine-Grained Long-Tailed Plant Taxonomy (ECCV 2026 · similarity 0.89)
- BioTrove: A Large Curated Image Dataset Enabling AI for Biodiversity (NeurIPS 2024 · similarity 0.86)
- INQUIRE: A Natural World Text-to-Image Retrieval Benchmark (NeurIPS 2024 · similarity 0.86)
- Prompt-CAM: Making Vision Transformers Interpretable for Fine-Grained Analysis (CVPR 2025 · similarity 0.86)
- DivShift: Exploring Domain-Specific Distribution Shifts in Large-Scale, Volunteer-Collected Biodiversity Datasets (AAAI 2025 · similarity 0.86)
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.
