Research map / Label-efficient and robust learning / Self-supervised and contrastive
CSP: Self-Supervised Contrastive Spatial Pre-Training for Geospatial-Visual Representations
Gengchen Mai, Ni Lao, Yutong He, Jiaming Song, 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 CSP-MC-BLD spatial contrastive pre-training | geospatial-visual classification · iNat2018, labeled training ratios 5%, 10%, 20%, and 100% | Top1 accuracy: Sup. Only (grid) to CSP-MC-BLD: 8.16 (0.01) -> 9.01 (0.02); 14.65 (0.03) -> 19.68 (0.05); 25.40 (0.05) -> 29.61 (0.03); | Improved |
| Add CSP-MC-BLD spatial contrastive pre-training | geospatial-visual classification · fMoW, labeled training ratios 5%, 10%, 20%, and 100% | Top1 accuracy: Sup. Only (grid) to CSP-MC-BLD: 67.01 (0.02) -> 67.47 (0.02); 68.91 (0.04) -> 69.23 (0.03); 70.20 (0.03) -> 70.66 (0.03) | Improved |
| Use the MC objective instead of NCE | CSP-BLD · iNat2018, all four training ratios | Top1 accuracy: 8.65 (0.02) -> 9.01 (0.02); 18.75 (0.12) -> 19.68 (0.05); 28.15 (0.07) -> 29.61 (0.03); 73.33 (0.01) -> 73.79 (0.02) | Improved |
| Use the MC objective instead of NCE | CSP-BLD · fMoW, all four training ratios | Top1 accuracy: 67.29 (0.03) -> 67.47 (0.02); 69.20 (0.03) -> 69.23 (0.03); 70.65 (0.02) -> 70.66 (0.03); 70.89 (0.04) -> 71.00 (0.02) | Improved |
| Add random negative location sampling (ablation) | CSP-MC-BLD · iNat2018 validation, 5% | Top1 accuracy: 8.63 -> 9.01 | Improved |
| Add SimCSE sampling (ablation) | CSP-MC-BLD · iNat2018 validation, 5% | Top1 accuracy: 8.40 -> 9.01 | Improved |
| Add random negative location and SimCSE sampling (ablation) | CSP-MC-B · iNat2018 validation, 5% | Top1 accuracy: 8.16 -> 9.01 | Improved |
| Set location embedding dimension to 256 (ablation) | CSP-MC-BLD · iNat2018 validation, 5%, 10%, and 20% | Top1 accuracy: d=256: 9.01, 19.68, 29.61; the paper reports these dimensions achieve the best performance at few-shot ratios | Improved |
| Set location embedding dimension to 1024 (ablation) | CSP-MC-BLD · iNat2018 validation, 100% | Top1 accuracy: 73.79 | Improved |
| Use ViT instead of InceptionV3 as the image encoder (ablation) | CSP-MC-BLD · iNat2018 validation, 5% | Top1 accuracy: 9.01 -> 20.78 | Improved |
Closest papers by meaning
- Geography-Aware Self-Supervised Learning (ICCV 2021 · similarity 0.91)
- SimIPU: Simple 2D Image and 3D Point Cloud Unsupervised Pre-training for Spatial-Aware Visual Representations (AAAI 2022 · similarity 0.88)
- Textual Supervision Enhances Geospatial Representations in Vision-Language Models (ICML 2026 · similarity 0.87)
- TorchSpatial: A Location Encoding Framework and Benchmark for Spatial Representation Learning (NeurIPS 2024 · similarity 0.86)
- TIPS: Text-Image Pretraining with Spatial awareness (ICLR 2025 · similarity 0.86)
- SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery (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.
