Research map / Label-efficient and robust learning / Self-supervised and contrastive
Seasonal Contrast: Unsupervised Pre-Training From Uncurated Remote Sensing Data
Oscar Mañas, Alexandre Lacoste, Xavier Giró-i-Nieto, David Vazquez, Pau Rodríguez
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 |
|---|---|---|---|
| sample pretraining locations around cities (Gaussian) instead of uniformly (ablation) | SeCo ResNet-18 (100k) · BigEarthNet linear probing 100% labels | mAP: 72.59 -> 75.52 | Improved |
| sample pretraining locations around cities (Gaussian) instead of uniformly (ablation) | SeCo ResNet-18 (100k) · BigEarthNet fine-tuning 100% labels | mAP: 85.75 -> 87.04 | Improved |
| sample pretraining locations around cities (Gaussian) instead of uniformly (ablation) | SeCo ResNet-18 (100k) · BigEarthNet linear probing 10% labels | mAP: 71.63 -> 74.67 | Improved |
| add temporal positive pairs to MoCo-v2 (ablation) | MoCo-v2 ResNet-18 (1M) · BigEarthNet linear probing 100% labels | mAP: 70.79 -> 73.60 | Improved |
| add temporal positive pairs to MoCo-v2 (ablation) | MoCo-v2 ResNet-18 · OSCD change detection | F1: 40.71 -> 40.12 | No clear effect |
| add multiple embedding sub-spaces (SeCo) to MoCo-v2+TP (ablation) | MoCo-v2+TP ResNet-18 (1M) · BigEarthNet linear probing 100% labels | mAP: 73.60 -> 77.00 | Improved |
| add multiple embedding sub-spaces (SeCo) to MoCo-v2+TP (ablation) | MoCo-v2+TP ResNet-18 · OSCD change detection | F1: 40.12 -> 46.94 | Improved |
| add multiple embedding sub-spaces (SeCo) to MoCo-v2+TP (ablation) | MoCo-v2+TP ResNet-18 · EuroSAT | accuracy: 89.51 -> 93.14 | Improved |
| scale pretraining set from 100k to 1M images (ablation) | SeCo ResNet-18 · BigEarthNet linear probing 100% labels | mAP: 75.52 -> 77.00 | Improved |
| scale pretraining set from 100k to 1M images (ablation) | SeCo ResNet-50 · BigEarthNet linear probing 100% labels | mAP: 79.13 -> 80.35 | Improved |
| replace ImageNet supervised pretraining with SeCo | ResNet-50 · BigEarthNet fine-tuning 100% labels | mAP: 86.74 -> 87.81 | Improved |
| replace ImageNet supervised pretraining with SeCo | ResNet-18 · EuroSAT | accuracy: 86.44 -> 93.14 | Improved |
Closest papers by meaning
- Rethinking Transformers Pre-training for Multi-Spectral Satellite Imagery (CVPR 2024 · similarity 0.85)
- Parameter Efficient Self-Supervised Geospatial Domain Adaptation (CVPR 2024 · similarity 0.85)
- Galileo: Learning Global & Local Features of Many Remote Sensing Modalities (ICML 2025 · similarity 0.85)
- SkySense: A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery (CVPR 2024 · similarity 0.84)
- Content-Adaptive Non-Local Convolution for Remote Sensing Pansharpening (CVPR 2024 · similarity 0.84)
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 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.
