Research map / Label-efficient and robust learning / Self-supervised and contrastive
Geography-Aware Self-Supervised Learning
Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, Stefano Ermon
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 geo-location classification pre-text task (ablation) | MoCo-V2 · fMoW single-image classification (frozen features) | accuracy: 60.69 -> 64.07 | Improved |
| add temporal positive pairs (ablation) | MoCo-V2 · fMoW single-image classification (frozen features) | accuracy: 60.69 -> 68.32 | Improved |
| add geo-location task on top of temporal positives (ablation) | MoCo-V2+TP · fMoW single-image classification (frozen features) | accuracy: 68.32 -> 66.33 | Hurt |
| add geo-location task and temporal positives, end-to-end finetuning (ablation) | MoCo-V2 · fMoW single-image classification (finetune) | accuracy: 64.34 -> 70.60 | Improved |
| pre-train MoCo-V2 on fMoW instead of ImageNet (ablation) | MoCo-V2 · fMoW single-image classification (finetune) | accuracy: 62.90 -> 64.34 | Improved |
| classify temporal data per area instead of single images (ablation) | MoCo-V2 · fMoW temporal classification (frozen features) | accuracy: 60.69 -> 68.64 | Improved |
| add temporal positives and geo-location task (ablation) | MoCo-V2 · xView object detection | AP: 15.45 -> 17.74 | Improved |
| add geo-location task alone (ablation) | MoCo-V2 · xView object detection | AP: 15.45 -> 15.63 | No clear effect |
| add temporal positives and geo-location task (ablation) | MoCo-V2 · SpaceNet segmentation | mIoU: 78.05 -> 78.51 | Improved |
| add geo-location task (ablation) | MoCo-V2 · NAIP land cover classification | top-1 accuracy: 55.18 -> 58.23 | Improved |
| add temporal positives to geo-location task (ablation) | MoCo-V2+Geo · NAIP land cover classification | top-1 accuracy: 58.23 -> 57.63 | Hurt |
| add geo-location task (ablation) | MoCo-V2 · GeoImageNet | top-1 accuracy: 38.51 -> 39.96 | Improved |
Closest papers by meaning
- CSP: Self-Supervised Contrastive Spatial Pre-Training for Geospatial-Visual Representations (ICML 2023 · similarity 0.91)
- Change-Aware Sampling and Contrastive Learning for Satellite Images (CVPR 2023 · similarity 0.90)
- Seasonal Contrast: Unsupervised Pre-Training From Uncurated Remote Sensing Data (ICCV 2021 · similarity 0.88)
- Sparse and Complete Latent Organization for Geospatial Semantic Segmentation (CVPR 2022 · similarity 0.87)
- Contextualized Spatio-Temporal Contrastive Learning With Self-Supervision (CVPR 2022 · similarity 0.87)
- Balanced Hierarchical Contrastive Learning with Decoupled Queries for Fine-grained Object Detection in Remote Sensing Images (CVPR 2026 · similarity 0.87)
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.
