The official code repository for the paper "Revisiting the Encoding of Satellite Image Time Series".
The preprint is under review.The codebase is still under construction and therefore is subject to further modifications.- The paper has been accepted to BMVC 2023 as an oral presentation.
- The model weights have been made available on Zenodo.
- The slides, poster, and accompanying video will be released after BMVC on a separate project page.
- I have been focusing on expanding this work to a journal paper and the code is subject to further modifications.
New SOTA Results on PASTIS Benchmark Dataset
| Model Name | mIoU | #Params (M) | FLOPs |
|---|---|---|---|
| U-TAE | 63.1 | 1.09 | 47G |
| TSViT | 65.4 | 2.16 | 558G |
| Exchanger+Unet | 66.8 | 8.08 | 300G |
| Exchanger+Mask2Former | 67.9 | 24.59 | 329G |
| Model Name | SQ | RQ | PQ | #Params (M) | FLOPs |
|---|---|---|---|---|---|
| UConvLSTM+PaPs | 80.2 | 43.9 | 35.6 | 2.50 | 55G |
| U-TAE+PaPs | 81.5 | 53.2 | 43.8 | 1.26 | 47G |
| Exchanger+Unet+PaPs | 80.3 | 58.9 | 47.8 | 9.99 | 301G |
| Exchanger+Mask2Former | 84.6 | 61.6 | 52.6 | 24.63 | 332G |
- The panoptic segmentation model Exchanger+Mask2Former has been trained by splitting the input into four 64x64 patches and stitch the prediction results together. Later on, I found this trick is crucial for replicating the results.
If you find our work or code useful in your research, please consider citing the following BibTex entry:
@article{cai2023rethinking,
title={Rethinking the Encoding of Satellite Image Time Series},
author={Cai, Xin and Bi, Yaxin and Nicholl, Peter and Sterritt, Roy},
journal={arXiv preprint arXiv:2305.02086},
year={2023}
}The codebase is built upon the following great work:
I would like to thank Zenodo for hosting the model weights and appreciate the constructive and insightful comments from BMVC reviewers.