We release data collected in the rugged Scottish highlands in extreme weather. The radar data we offer to the community are accompanied by GPS/INS reference - to further stimulate research in radar place recognition

Data

We release data from four areas of the Ardverikie Estate. These feature distinct landscape (therefore typical radar returns) as well as driving conditions. In total we drive a diverse set of four routes over 11 forays, totalling approximately 154km of rugged driving.

Please follow the (radar, gps) links below to download the data.

Bellmouth	Hydro
2021-11-25-12-01-20 (radar, images, gps)	2021-11-27-14-37-20 (radar, images, gps)
2021-11-25-12-31-19 (radar, images, gps)	2021-11-27-15-24-02 (radar, images, gps)
2021-11-26-15-35-34 (radar, images, gps)	2021-11-27-16-03-26 (radar, images, gps)
2021-11-26-16-12-01 (radar, gps)
Maree	Two Lochs
2021-11-28-15-54-55 (radar, images, gps)	2021-11-29-11-40-37 (radar, images, gps)
2021-11-28-16-43-37 (radar, images, gps)	2021-11-29-12-19-16 (radar, images, gps)

Software Development Kit

A software development kit is available at github.com/mttgdd/oord-dataset. This allows quick setup of place recognition experiments and ground truth performance evaluation, as below.

We start by specifying the configuration of our experiment in config/alexnet.yaml.

tar_dir: /Volumes/saxdata2/data
out_dir: /home/mattgadd/experiments/oord/test
data_dir: /home/mattgadd/data/oord

seed: 42
batch_size: 16
device: cuda:0
downsample: 5
pos_gps: 25.0
neg_gps: 50.0

min_bin: 60
max_bin: 3768
cartesian: True
cart_res: 1.2717
cart_pw: 224
channels: 3
fft: False

gps_match_tolerance: 25.0

We then write a simple wrapper jobs/test_alexnet.py around a common test routine in src/test.py

from config.config import get_cfg_impl
from src.test import test

import torch

cfg = get_cfg_impl('config/alexnet.yaml')

net = torch.hub.load('pytorch/vision:v0.10.0', 'alexnet', pretrained=True)
net.eval()

test(cfg, net, None, 'alexnet', sensor='radar')

Running the job with python -m jobs.test_alexnet, the embedding distance matrix are exported as pngs and we can also generate all ground truth matrices as pngs with:

python -m tex.figs_gps

This gives us the following example outputs at /home/mattgadd/experiments/oord/test/alexnet_2021-11-25-12-31-19_2021-11-25-12-01-20.png and /home/mattgadd/experiments/oord/test/2021-11-25-12-31-19_2021-11-25-12-01-20-positives.png:

The recall@n metrics are run as part of the call to src/test.py above and are written to e.g. /home/mattgadd/experiments/oord/test/alexnet_2021-11-25-12-31-19_2021-11-25-12-01-20.csv with the following example contents.

,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49
0,0.7491190979563073,0.8463706835799859,0.8787878787878788,0.9062720225510923,0.9238900634249472,0.9330514446793516,0.9400986610288936,0.9443269908386187,0.9471458773784355,0.952783650458069,0.959830866807611,0.9626497533474278,0.9640591966173362,0.9654686398872445,0.9704016913319239,0.9760394644115574,0.9774489076814659,0.9795630725863284,0.9802677942212826,0.9809725158562368,0.9830866807610994,0.9830866807610994,0.9859055673009162,0.9859055673009162,0.9859055673009162,0.9866102889358703,0.9880197322057788,0.9880197322057788,0.9880197322057788,0.9894291754756871,0.9894291754756871,0.9894291754756871,0.9908386187455955,0.9908386187455955,0.9908386187455955,0.9908386187455955,0.9908386187455955,0.9908386187455955,0.9915433403805497,0.9915433403805497,0.9915433403805497,0.9922480620155039,0.992952783650458,0.9943622269203665,0.9943622269203665,0.9943622269203665,0.9950669485553206,0.9950669485553206,0.9950669485553206

Reading this, we have 74.91% recall@1 for alexnet.

Trained model checkpoints

For the neural network approaches explored in the paper, we release 4 saved checkpoints at huggingface.co/mttgdd/oord-models. Example use is as below:

from src.resnet18_netvlad import Net

from huggingface_hub import hf_hub_download
import torch

net = Net()
weights = hf_hub_download(
      repo_id="mttgdd/oord_models", filename="resnet18_netvlad_oord-2021-11-25-12-31-19.pth")
net.load_state_dict(torch.load(weights))

Citation

Please cite the following paper when using the dataset in your research.

OORD: The Oxford Offroad Radar Dataset
M. Gadd, D. De Martini, O. Bartlett, P. Murcutt, M. Towlson, M. Widojo, V. Muşat, L. Robinson, E. Panagiotaki, G. Pramatarov, M. A. Kühn, L. Marchegiani, P. Newman, L. Kunze
arXiv preprint arXiv:2403.02845, 2024
[Bibtex]

@article{gadd2024oord,
title={{OORD: The Oxford Offroad Radar Dataset}},
author={Gadd, Matthew and De Martini, Daniele and Bartlett, Oliver and Murcutt, Paul and Towlson, Matt and Widojo, Matthew and Mu\cb{s}at, Valentina and Robinson, Luke and Panagiotaki, Efimia and Pramatarov, Georgi and K"uhn, Marc Alexander and Marchegiani, Letizia and Newman, Paul and Kunze, Lars},
journal={arXiv preprint arXiv:2403.02845},
year={2024}
}