---
license: cc-by-sa-4.0
tags:
- visual-place-recognition
- robotics
- season-change
- place-recognition
size_categories:
- 10K<n<100K
task_categories:
- image-to-image
configs:
- config_name: default
  data_files:
  - split: summer2012
    path: summer2012/train-*.parquet
  - split: summer2015
    path: summer2015/train-*.parquet
  - split: winter
    path: winter/train-*.parquet
dataset_info:
  features:
  - name: filename
    dtype: string
  - name: image
    dtype: image
  splits:
  - name: summer2012
    num_examples: 6915
  - name: summer2015
    num_examples: 5392
  - name: winter
    num_examples: 30790
---
# Freiburg Across Seasons Dataset

Image sequences recorded in the city of Freiburg across three years for benchmarking long-term, season-invariant Visual Place Recognition (VPR). The recordings were captured by Naseer, Spinello, Stachniss and Burgard at the University of Freiburg with a forward-facing Bumblebee stereo camera mounted on a car. The dataset preparation, ground-truth annotation and original distribution are by the authors. This Hugging Face dataset card repackages the data for convenient consumption via the `datasets` library; the underlying recordings, processing, and ground truth are theirs.

## Dataset Details

- **Source recordings & ground truth:** Tayyab Naseer, Luciano Spinello, Cyrill Stachniss, Wolfram Burgard, University of Freiburg
- **Papers:**
  - Naseer, Burgard, Stachniss. *Robust Visual Localization Across Seasons.* IEEE Transactions on Robotics (TRO), 2018.
  - Naseer, Spinello, Burgard, Stachniss. *Robust Visual Robot Localization Across Seasons using Network Flows.* AAAI, 2014.
- **Original distribution:** http://aisdatasets.informatik.uni-freiburg.de/freiburg_across_seasons/
- **License:** [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/) (per the authors' original release)

## Recording Setup

The dataset preparation follows Naseer et al. (2014/2018):

- Recordings were captured with a forward-facing Bumblebee stereo camera mounted on a car. In summer the camera was mounted outside the car, in winter inside.
- All images are 1024×768 (width×height), JPEG-compressed.
- No preprocessing was applied to the recorded images.
- All images are geo-tagged with corresponding GPS positions (recorded with an inexpensive GPS sensor).
- Three sequences were recorded:
  - **Localization-1** (Summer 2012): 6,915 images at 1 Hz, covering 10 km.
  - **Mapping** (Winter 2012): 30,790 images at 4 Hz, covering 50 km.
  - **Localization-2** (Summer 2015): 5,392 images.
- Ground truth was created by hand-labelling matches after manual visual inspection. Localization-1 has 3,656 ground-truth-matched images to the Mapping sequence, Localization-2 has 4,477. In total, 8,133 ground-truth matchings are provided.

## What this Hugging Face version contains

This Hugging Face card distributes a subset of the original Freiburg Across Seasons release:

- **Camera 0 only.** Only the left camera of the Bumblebee stereo pair is included. The right-camera stream is not part of this distribution, so stereo depth extraction is not possible from this version. For the full stereo data, refer to the [original release](http://aisdatasets.informatik.uni-freiburg.de/freiburg_across_seasons/).
- **Ground-truth matchings only as annotations.** The manually annotated query→reference correspondences from Naseer et al. are provided as Parquet files in `matchings/`. No additional labels (semantic, depth, pose) are included.
## Traversals

| Folder | Sequence | Season | Images |
|---|---|---|---|
| `summer_icra_dataset/` | Localization-1 | Summer 2012 | 6,915 |
| `summer2015_icra_dataset/` | Localization-2 | Summer 2015 | 5,392 |
| `winter_icra_dataset/` | Mapping | Winter 2012 | 30,790 |

## Ground-truth matchings

Located in `matchings/`:

| File | Query traversal | Reference traversal | Pairs |
|---|---|---|---|
| `summer2012_winter_matchings.parquet` | summer2012 | winter | 4,037 |
| `summer2015_winter_matchings.parquet` | summer2015 | winter | 4,739 |

Each Parquet file has two integer columns, `<season>_idx` and `winter_idx`, which are zero-based indices into the corresponding image folder sorted lexicographically by filename.

## Loading example

```python
import pandas as pd
from PIL import Image
import os

df = pd.read_parquet("matchings/summer2015_winter_matchings.parquet")
summer_imgs = sorted(os.listdir("summer2015_icra_dataset/"))
winter_imgs = sorted(os.listdir("winter_icra_dataset/"))

query = Image.open(f"summer2015_icra_dataset/{summer_imgs[df.iloc[0]['summer2015_idx']]}")
match = Image.open(f"winter_icra_dataset/{winter_imgs[df.iloc[0]['winter_idx']]}")
```

## License

Released under [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/), matching the original distribution by Naseer et al. Attribution and ShareAlike conditions apply.

## Citation
Please cite these publications if you use this dataset.
The papers can be found here: [Robust Visual Localization Across Seasons](https://ieeexplore.ieee.org/document/8269404) and [Robust Visual Robot Localization Across Seasons using Network Flows](https://ojs.aaai.org/index.php/AAAI/article/view/9057).


```bibtex
@article{naseerRobustVisualLocalization2018,
    title   = {Robust Visual Localization Across Seasons},
    author  = {Naseer, Tayyab and Burgard, Wolfram and Stachniss, Cyrill},
    journal = {IEEE Transactions on Robotics},
    year    = {2018}
}

@inproceedings{naseerRobustVisualRobot2014,
    title     = {Robust Visual Robot Localization Across Seasons using Network Flows},
    author    = {Naseer, Tayyab and Spinello, Luciano and Burgard, Wolfram and Stachniss, Cyrill},
    booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
    address   = {Quebec, Canada},
    year      = {2014}
}
```