CLEAR-Net for SEN12MS-CR
This repository hosts the best CLEAR-Net checkpoint for cloud removal experiments on SEN12MS-CR style SAR and cloudy optical inputs.
The implementation, demo server, and local inference launcher are maintained in the GitHub repository:
GitHub: https://github.com/smturtle2/cr-project
Checkpoint
| File | Description |
|---|---|
best.pt |
PyTorch checkpoint for CLEAR_Net(return_decomposition=True) |
The checkpoint is stored in the project training format. The inference code expects a PyTorch checkpoint containing a model state dict.
Checkpoint metadata:
| Field | Value |
|---|---|
| Epoch | 87 |
| Global step | 291363 |
| SHA256 | 6b6a001c29d95e9edabd415e5856b83f41bea63c76ef05c5e85312147c697eb4 |
Test Metrics
Evaluation on the project test split:
| Metric | Value |
|---|---|
| MAE | 0.027414 |
| PSNR | 28.4613 |
| SSIM | 0.893636 |
| SAM | 8.2448 |
| Loss | 0.313022 |
Usage
Clone the project code:
git clone https://github.com/smturtle2/cr-project.git
cd cr-project
Download the checkpoint:
hf download Hermanni/clear-net-sen12mscr best.pt --local-dir .
Run the local demo:
./inference.sh ./best.pt
If the SEN12MS-CR cache is stored outside the project default path, pass it through the demo arguments:
./inference.sh ./best.pt --dataset-root /path/to/sen12mscr_cache
The demo opens a local CLEAR-Net scene inference server and uses the checkpoint from this repository.
Intended Use
This checkpoint is intended for research and demonstration of optical cloud removal using SAR-guided restoration on SEN12MS-CR style data.
Limitations
The model card reports results from the project evaluation pipeline. Performance may differ with different preprocessing, dataset versions, sensor products, or scene tiling settings.