Research map / Detection and segmentation / CNN features
Harnessing Massive Satellite Imagery with Efficient Masked Image Modeling
Fengxiang Wang, Hongzhen Wang, Di Wang, Zonghao Guo, Zhenyu Zhong, Long Lan, Wenjing Yang, Jing Zhang
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 |
|---|---|---|---|
| replace MAE with SelectiveMAE (HOG selection, PSTS, partial reconstruction) | ViT-B pre-training · OpticalRS-13M pre-training | data throughput per minute: 264k -> 556k | Improved |
| replace MAE with SelectiveMAE | ViT-B, 1M images · Pre-training efficiency | training time (h): 51 -> 24 | Improved |
| replace MAE with SelectiveMAE | ViT-L, 1M images · Pre-training efficiency | training time (h): 57 -> 25 | Improved |
| replace MAE with SelectiveMAE | ViT-B, 1M images · Pre-training efficiency | GPU memory (MB): 17628 -> 9570 | Improved |
| replace MAE with SelectiveMAE | ViT-B · AID (TR=20%) | overall accuracy: 96.58 -> 96.90 | Improved |
| replace MAE with SelectiveMAE | ViT-B · SpaceNetv1 | mF1: 79.41 -> 79.50 | No clear effect |
| scale SelectiveMAE from ViT-B to ViT-L (ablation) | SelectiveMAE · DIOR-R | mAP50: 67.69 -> 71.75 | Improved |
| add HOG selection instead of Swin-B feature clustering (ablation) | SelectiveMAE · AID (TR=20%, MillionAID) | overall accuracy: 93.21 -> 93.17 | No clear effect |
| select tokens by HOG instead of Adamae sampler (ablation) | SelectiveMAE · AID (TR=20%, MillionAID) | overall accuracy: 88.78 -> 93.17 | Improved |
| reconstruct HOG-selected patches instead of random patches (ablation) | partial reconstruction · AID (TR=20%, MillionAID) | overall accuracy: 94.1 -> 94.8 | Improved |
| use near-far-random selection in PSTS instead of far-near-random (ablation) | SelectiveMAE · AID (TR=20%, MillionAID) | overall accuracy: 90.12 -> 93.17 | Improved |
| pre-train on OpticalRS-13M subset instead of MillionAID (ablation) | MAE ViT-B · AID (TR=20%, 1M images) | overall accuracy: 94.92 -> 96.26 | Improved |
Closest papers by meaning
- S2MAE: A Spatial-Spectral Pretraining Foundation Model for Spectral Remote Sensing Data (CVPR 2024 · similarity 0.87)
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.87)
- Accounting for Spatial Variability with the Histogram of Oriented Gradients Based Masking Improves Performance of Masked Autoencoder over Hyperspectral Satellite Imagery (Student Abstract) (AAAI 2025 · similarity 0.87)
- Learning De-Biased Representations for Remote-Sensing Imagery (NeurIPS 2024 · similarity 0.86)
- SkySense: A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery (CVPR 2024 · similarity 0.86)
- Rethinking Transformers Pre-training for Multi-Spectral Satellite Imagery (CVPR 2024 · similarity 0.85)
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.
