README.md

---
license: mit
task_categories:
- depth-estimation
- keypoint-detection
- image-feature-extraction
pretty_name: AerialExtreMatch Train
viewer: false
tags:
- image
---
# AerialExtreMatch — Train Dataset

[Code](https://github.com/Xecades/AerialExtreMatch) | [Project Page](https://xecades.github.io/AerialExtreMatch/) | Paper (WIP)

This repo contains the **training** set for our paper *AerialExtreMatch: A Benchmark for Extreme-View Image Matching and Localization*. We also provide [**benchmark**](https://huggingface.co/datasets/Xecades/AerialExtreMatch-Benchmark) and [**localization**](https://huggingface.co/datasets/Xecades/AerialExtreMatch-Localization) datasets.

## Usage

Simply clone this repository and unzip the dataset files.

```bash
git clone git@hf.co:datasets/Xecades/AerialExtreMatch-Train
cd AerialExtreMatch-Train
unzip "*.zip"
rm -rf *.zip
rm -rf .git
```

## Dataset Structure

After unpacking each .zip file:

<pre>
.
├── scene_info: *.npz <i>(e.g. Tokyo3_100@-98@-185@67@0@90.npz)</i>
├── geo_info: *.txt <i>(e.g. Tokyo_3.txt, see below)</i>
└── [city_id] <i>(e.g. Tokyo3)</i>
    ├── computed_pairs.txt
    ├── overlap_5_30.txt
    ├── overlap_30_50.txt
    ├── overlap_50_plus.txt
    ├── query
    │   ├── intrinsic.txt <i>(see below)</i>
    │   ├── poses.txt <i>(see below)</i>
    │   ├── depth: *.exr
    │   └── rgb: *.jpg
    └── reference
        ├── intrinsic.txt <i>(see below)</i>
        ├── poses.txt <i>(see below)</i>
        ├── depth: *.exr
        └── rgb: *.jpg
</pre>

 - Keys of `.npz` files in `scene_info`: `['poses', 'intrinsics', 'depth', 'rgb', 'overlap', 'pair']`.
 - Format of `.txt` files in `geo_info`: `name lat1 lon1 lat2 lon2 alt`.
 - Format of `intrinsic.txt`: `name PINHOLE w h fx fy cx cy`.
 - Format of `poses.txt`: `name qw qx qy qz x y z`. (w2c)