Research map / Label-efficient and robust learning / Self-supervised and contrastive
Focus on the Positives: Self-Supervised Learning for Biodiversity Monitoring
Omiros Pantazis, Gabriel J. Brostow, Kate E. Jones, Oisin Mac Aodha
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 |
|---|---|---|---|
| select positives by sequence (same location, within 5 s) instead of augmentation | SimCLR · CCT20 (1% labels) | top-1 accuracy: 65.75 -> 67.60 | Improved |
| select positives by context similarity instead of augmentation | SimCLR · CCT20 (1% labels) | top-1 accuracy: 65.75 -> 68.67 | Improved |
| select positives by sequence instead of augmentation | Triplet · CCT20 (1% labels) | top-1 accuracy: 59.13 -> 66.30 | Improved |
| select positives by context instead of augmentation | SimSiam · CCT20 (1% labels) | top-1 accuracy: 56.64 -> 59.93 | Improved |
| select positives by context instead of augmentation | SimCLR · MMCT (1% labels, GT boxes) | top-1 accuracy: 54.93 -> 64.01 | Improved |
| select positives by context instead of augmentation | SimSiam · MMCT (1% labels, GT boxes) | top-1 accuracy: 48.78 -> 61.66 | Improved |
| use context positives on a satellite dataset | SimCLR · FMoW subset (1% labels) | top-1 accuracy: 45.60 -> 45.93 | Improved |
| use sequence positives instead of standard on FMoW | SimCLR · FMoW subset (100% labels) | top-1 accuracy: 52.16 -> 50.85 | Hurt |
| initialize backbone with ImageNet instead of random weights (ablation) | SimCLR Standard · CCT20 (1% labels) | top-1 accuracy: 53.29 -> 65.75 | Improved |
| use context positives under random init (ablation) | SimCLR · CCT20 (1% labels, random init) | top-1 accuracy: 53.29 -> 58.86 | Improved |
| increase capacity to ResNet50 at 224px (ablation) | SimCLR Standard · CCT20 (1% labels) | top-1 accuracy: 65.75 -> 68.59 | Improved |
| fine-tune end-to-end after context SSL instead of linear eval (ablation) | SimCLR Seq. Pos. · CCT20 (1% labels) | top-1 accuracy: 67.60 -> 72.56 | Improved |
Closest papers by meaning
- Benchmarking Representation Learning for Natural World Image Collections (CVPR 2021 · similarity 0.91)
- Lessons and Open Questions from a Unified Study of Camera-Trap Species Recognition Over Time (ECCV 2026 · similarity 0.87)
- Dynamic Curriculum Learning for Great Ape Detection in the Wild (IJCV 2023 · similarity 0.87)
- Contrastive ground-level image and remote sensing pre-training improves representation learning for natural world imagery (ECCV 2024 · similarity 0.86)
- BioTrove: A Large Curated Image Dataset Enabling AI for Biodiversity (NeurIPS 2024 · similarity 0.84)
- DivShift: Exploring Domain-Specific Distribution Shifts in Large-Scale, Volunteer-Collected Biodiversity Datasets (AAAI 2025 · similarity 0.84)
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.
