Datasets:
metadata
license: cc-by-4.0
task_categories:
- image-to-image
tags:
- vpr
- nordland
- seasonal-changes
- place-recognition
pretty_name: Nordland Dataset (Olid et al. Partition)
size_categories:
- 10K<n<100K
dataset_info:
features:
- name: image
dtype: image
- name: season
dtype:
class_label:
names:
'0': fall
'1': spring
'2': summer
'3': winter
- name: frame_id
dtype: int32
- name: section
dtype:
class_label:
names:
'0': train_section1
'1': train_section2
'2': test_section1
'3': test_section2
'4': test_section3
splits:
- name: train
num_bytes: 161117439484
num_examples: 98280
- name: test
num_bytes: 22934069019
num_examples: 13800
download_size: 184051508503
dataset_size: 184051508503
configs:
- config_name: default
data_files:
- split: train
path: data/train-*.parquet
- split: test
path: data/test-*.parquet
Nordland Dataset (Olid et al. Partition)
Pre-processed and partitioned version of the Nordland Dataset by Olid et al.
Figure reproduced from Olid et al. (2018), Single-View Place Recognition under Seasonal Changes, Fig. 3. © The original authors.
Dataset Details
- Source video: Norwegian Broadcasting Corporation (NRK), Nordlandsbanen: minute by minute, season by season
- Pre-processing & partition: Daniel Olid, Jose M. Facil, Javier Civera, Universidad de Zaragoza
- Paper: Single-View Place Recognition under Seasonal Changes, PPNIV Workshop at IROS 2018
Important Things to Know
- One frame per second was extracted from the original Nordland videos. Tunnels and train stops were filtered out, leaving 28,865 images per season.
- Each image is somewhere between 1m and 25m away from the previous and the next one.
- The test split consists of 3 sections of 1,150 images each.
- The train split consists of 2 sections, the first with 12,837 images and the second with 11,733.
- 200 images were discarded between each section so that train and test sections represent different geographic places.
Dataset Structure
Each sample contains:
image— PNG frame extracted from the Nordland train-journey videoseason— one offall,spring,summer,winterframe_id— global frame index (same index across seasons = same geographic location)section— one oftrain_section1,train_section2,test_section1,test_section2,test_section3
Sections
| Section | frame_id range |
Images per season |
|---|---|---|
train_section1 |
0 – 12836 | 12,837 |
train_section2 |
12837 – 24569 | 11,733 |
test_section1 |
0 – 1149 | 1,150 |
test_section2 |
1150 – 2299 | 1,150 |
test_section3 |
2300 – 3449 | 1,150 |
Usage
from datasets import load_dataset, Value
ds = load_dataset("fansel/nordland")
# season and section are stored as ClassLabel
# Cast them to plain strings once for natural filtering:
ds = (ds
.cast_column("season", Value("string"))
.cast_column("section", Value("string")))
# Access splits
train, test = ds["train"], ds["test"]
# Filter by season
winter_test = test.filter(lambda x: x["season"] == "winter")
# Filter train section 1, all seasons
train_section1 = train.filter(lambda x: x["section"] == "train_section1")
# Filter test section 2, season fall
test_s2_fall = test.filter(
lambda x: x["section"] == "test_section2" and x["season"] == "fall"
)
# Cross-season matching: same frame_id across seasons
same_location = test.filter(lambda x: x["frame_id"] == 42)
for sample in same_location:
print(sample["season"], sample["section"])
Citation
Please cite this publication if you use this partitioned version of the dataset.
@article{olidSingleViewPlaceRecognition2018,
title = {Single-{View} {Place} {Recognition} under {Seasonal} {Changes}},
volume = {abs/1808.06516},
url = {https://arxiv.org/abs/1808.06516},
journal = {ArXiv preprint},
author = {Olid, Daniel and Fácil, José M. and Civera, Javier},
year = {2018}
}