ECCV 2026 submission
SeasonStereo
Robust dense stereo matching for multi-date satellite imagery via generative AI.
Abstract
Learning Dense Stereo Across Dates
Accurate 3D reconstruction from satellite imagery typically relies on near-simultaneous stereo pairs, limiting its applicability to diachronic settings where multi-date images exhibit varying seasonal and illumination conditions. Training dense stereo matching models robust to appearance changes is a long-standing challenge, as aligned multi-date imagery and ground-truth geometry are costly to obtain at scale. We propose SeasonStereo, a scalable framework that addresses disparity estimation from diachronic satellite images by training on synthetic image pairs with controlled seasonal appearance variation, while leveraging zero-shot geometric priors from foundation models. SeasonStereo matches the disparity accuracy of state-of-the-art LiDAR-supervised models, while producing sharper geometric details without requiring real aligned multi-date training products or LiDAR-derived labels. As a result, SeasonStereo offers a practical path toward large-scale 3D reconstruction from heterogeneous satellite images with reduced supervision cost.
Overview
Stereo Under Seasonal Change
SeasonStereo starts from a conservative set of reliable synchronic satellite stereo pairs, expands them into seasonal variants, and fine-tunes a strong stereo foundation model so dense matching remains stable when the two views come from different dates.
Synthetic Data
Seasonal Appearance, Preserved Geometry
Each selected crop is edited into four seasonal variants. The release page shows one Omaha crop and its generated spring, summer, autumn, and winter appearances.
Interactive Mesh
Geometry Viewer
The viewer uses real output products for the Omaha diachronic pair
OMA_084_042_RGB-OMA_084_016_RGB. Disparity surfaces show
image-space matching structure, while DSM surfaces show metric height
relative to each DSM's local minimum.
Model
Release
Reproducibility Package
Training Pipeline
SeasonStereo fine-tuning, evaluation scripts, loss functions, configs, and the MonSter++ wrapper used in the experiments.
Preprocessing
Rectification, semantic-mask generation, similarity scoring, pair selection, and pseudo-ground-truth disparity generation.
Data And Weights
Validation and test splits, cropped real tiles and their seasonal variants, semantic masks, a single-AOI sample of the training split, and the trained checkpoint. The full training split is not distributed; it is regenerated with the preprocessing pipeline.
Paper
Citation
@inproceedings{seasonstereo2026,
title = {SeasonStereo: Robust Dense Stereo Matching for Multi-Date Satellite Imagery via Generative AI},
author = {Authors},
booktitle = {Venue},
year = {2026}
}