Alvaritox/seasonstereo-data

SeasonStereo — Data & Checkpoints

Diachronic stereo matching for multi-date satellite imagery.

This repository hosts the datasets released with the SeasonStereo paper. The training/evaluation code lives in a separate GitHub repository: SeasonStereo. See that repo's README.md for installation, training, and evaluation instructions — this page only documents the data itself.

Links

• Code
• Project page
• Paper
• Checkpoint

Contents

Folder	Size	Description
Train-Track3-cropped/	~4.8 GB	Real cropped RGB tiles (Track3), .tif, organized by AOI
Train-Track3-cropped-synthetic/	~19 GB	Seasonal synthetic variants of the tiles above (.png, 4 seasons per tile)
water_segmentation/, tree_segmentation/, building_segmentation/	~1 GB total	Semantic masks + probability maps for the cropped tiles
diachronic-stereo-synthetic/	~11.5 GB (train 1 AOI) + ~50 GB (val/test/other)	Main stereo training/validation/test set: rectified pairs, disparity, masks, homographies

Train-Track3-cropped/ and Train-Track3-cropped-synthetic/

Train-Track3-cropped/
├── Track3-RGB-1/<AOI, e.g. JAX_204>/<AOI>_<tile_id>_RGB.tif
└── Track3-RGB-2/<AOI, e.g. OMA_367>/<AOI>_<tile_id>_RGB.tif

Train-Track3-cropped-synthetic/
├── Track3-RGB-1/<AOI>/<AOI>_<tile_id>_RGB_{SPRING,SUMMER,AUTUMN,WINTER}.png
└── Track3-RGB-2/<AOI>/<AOI>_<tile_id>_RGB_{SPRING,SUMMER,AUTUMN,WINTER}.png

Same AOI/tile layout as the real tiles; each real .tif has four seasonal synthetic .png counterparts generated for training.

Segmentation masks

segmentation_masks/{water,tree,building}_segmentation/
├── masks/<tile>_mask.png        # binary/class segmentation mask
└── probs/<tile>_{water,tree,building}_prob.png   # per-pixel probability map

diachronic-stereo-synthetic/

The main dataset used to train and evaluate the stereo model.

diachronic-stereo-synthetic/
├── train/
│   ├── L/, R/                       # rectified left/right image pairs, .npy, float32, HxWx3
│   ├── disparity/                   # pseudo-GT / reference disparity, .iio (numpy array format)
│   ├── masks/, water_masks/, tree_masks/, building_masks/   # each split into L/ and R/, .npy, uint8, HxW
│   └── homography/                  # rectification homographies, .npz with keys: Hleft, Hright, out_shape
├── val/                              # same structure as train/
├── test/
│   ├── JAX/       # test_jax, test_jax_dsm_and_rpc (stereo pairs + DSM/RPC metadata)
│   ├── OMA/       # test_omaha_synchronic, test_omaha_diachronic, generated_matching_dataset_v2, all_oma_and_jax_dsm_and_rpc
│   └── IARPA/     # test_buenos_aires, test_buenos_aires_dsm_and_rpc
├── synchronic_only/                  # L/, R/, disparity/, homography/ — synchronic-only real pairs
├── disparity_maps/
│   ├── GT/                           # ground-truth disparity (LiDAR-derived)
│   └── monster++/                    # MonSter++ pseudo-label disparity used for supervision
├── experiments/
│   ├── all.csv                       # full training pair list
│   ├── sub1_synchronic.csv           # real synchronic pairs only
│   ├── sub2_synchronic.csv           # real + synthetic synchronic pairs
│   ├── sub3_syndiach.csv             # real/synthetic + synchronic/diachronic pairs
│   └── sub_val.csv                   # validation pairs (105 of 109 original, see note below)
├── intersected_pairs.csv
├── train_aois.csv / val_aois.csv / test_aois.csv
└── description_exps.txt              # original notes on subsets/experiments (see below)

Reading the array files

import numpy as np

# .npy files (L, R, masks, etc.)
arr = np.load("train/L/<file>.npy")

# .iio files are numpy arrays with a non-standard extension — same reader works
with open("train/disparity/<file>.iio", "rb") as f:
    disp = np.lib.format.read_array(f)

# homography .npz
h = np.load("train/homography/<file>.npz")
Hleft, Hright, out_shape = h["Hleft"], h["Hright"], h["out_shape"]

Checkpoint

checkpoints/season-stereo-final.pth — final SeasonStereo model weights (experiment: disparity + photometric + aware smoothness losses, supervised with MonSter++ pseudo-labels). Load with the training/eval code in the SeasonStereo repo.

Local layout for use with the code repo

To use this data directly with the SeasonStereo codebase, place it as:

data/
  diachronic-stereo-synthetic/
  synchronic_only/
  Train-Track3-cropped/
  Train-Track3-cropped-synthetic/
  water_segmentation/
  tree_segmentation/
  building_segmentation/

checkpoints/
  season-stereo-final.pth

Citation

@inproceedings{diaz-laureano2026seasonstereo,
  title     = {SeasonStereo: Robust Dense Stereo Matching for Multi-Date Satellite Imagery via Generative AI},
  author    = {D{\'\i}az-Laureano, {\'A}lvaro and Mar{\'\i}, Roger and Masquil, El{\'\i}as and Arias, Pablo and Facciolo, Gabriele},
  booktitle = {Proceedings of the TerraBytes II Workshop at the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

Acknowledgments

The non-synthetic satellite images used in this project originate from the 2019 IEEE GRSS Data Fusion Contest (DFC2019), Track 3. The original WorldView-3 imagery is provided courtesy of DigitalGlobe (now Maxar). The DFC2019 dataset was organized by the IEEE GRSS Image Analysis and Data Fusion Technical Committee (IADF TC), the Johns Hopkins University Applied Physics Laboratory (JHU/APL), and the Intelligence Advanced Research Projects Activity (IARPA). All non-synthetic images and their copyright remain the property of their respective owners.

We are grateful to the DFC2019 organizers and data providers for making this large-scale satellite stereo dataset publicly available and enabling further research in 3D reconstruction from satellite imagery.

If you use our dataset, please also cite the original DFC2019 references:

@inproceedings{bosch2019semantic,
  title     = {Semantic Stereo for Incidental Satellite Images},
  author    = {Bosch, Marc and Foster, Kevin and Christie, Gordon and Wang, Sean and Hager, Gregory D. and Brown, Myron},
  booktitle = {2019 IEEE Winter Conference on Applications of Computer Vision (WACV)},
  pages     = {1524--1532},
  year      = {2019},
  organization = {IEEE}
}
 
@article{le_saux2019dfc,
  title   = {2019 IEEE GRSS Data Fusion Contest: Large-Scale Semantic 3D Reconstruction},
  author  = {Le Saux, Bertrand and Yokoya, Naoto and H{\"a}nsch, Ronny and Brown, Myron},
  journal = {IEEE Geoscience and Remote Sensing Magazine},
  volume  = {7},
  number  = {4},
  pages   = {33--36},
  year    = {2019}
}