Source-linked AI summary
YILDIZ-VPR: A Novel Dataset with Dense Coverage Under Diverse Environmental Conditions for Visual Place Recognition
Serdar Yildiz, Abbas Memiş, Songül Varli
TL;DR
Visual place recognition needs datasets combining dense spatial coverage with repeated observations across changing environmental conditions. YILDIZ-VPR addresses this gap with a pedestrian-level campus dataset, offering dense GPS-linked imagery and varied outdoor observations for image-based and temporal VPR. The resulting resource supports investigation of image-based, sequential, day-to-night, and multimodal visual place recognition.
Problem
VPR lacks sufficient datasets combining dense spatial coverage with repeated observations across varied times, seasons, weather, and viewpoints.
Method
YILDIZ-VPR uses repeated pedestrian-level walking traversals with synchronized GPS, filtered location annotations, and auxiliary sensor metadata.
Results
YILDIZ-VPR contains 370,464 training images, 2,377 daytime queries, and 1,902 nighttime queries across diverse outdoor scenes with GPS coordinates.
Takeaways & Limitations
The dataset offers a resource for investigating image-based, sequential, day-to-night, and multimodal visual place recognition.
Takeaways & Limitations
The dataset excludes indoor areas.
Abstract
from arXiv · showhide
Visual Place Recognition (VPR) aims to recognize the location of a query image by comparing it with a set of geo-referenced images. Although many datasets have been proposed for VPR, collecting dense and diverse visual data from pedestrian-level viewpoints is still an important need. In this paper, we introduce YILDIZ-VPR, a visual geo-localization dataset collected through repeated walking traversals on the Davutpasa campus of Yildiz Technical University. The dataset includes outdoor scenes captured at different times of day, seasons, and weather conditions. It contains a wide range of visual content, including historical buildings, modern structures, roads, green areas, and wooded regions. Each video was recorded with a GoPro 9 camera and synchronized with GPS sensor data to provide location labels for the extracted frames. In addition to GPS coordinates, the dataset also includes auxiliary sensor information such as gyroscope, speed, and temperature data. With its dense coverage and long-term visual variability, YILDIZ-VPR provides a useful resource for studying image-based and temporal visual place recognition under realistic outdoor conditions.
1. Introduction
Visual Place Recognition matches query images with geo-referenced gallery images to support localization, but remains challenging under environmental variation and unreliable GPS. YILDIZ-VPR addresses these needs with dense, repeated pedestrian-level traversals across diverse outdoor scenes on the Davutpasa campus.
- VPR motivation: VPR identifies locations by matching query images against geo-referenced gallery images and supports applications including navigation, robotics, augmented reality, and visual search.It is commonly formulated as an image retrieval problem using visual descriptors.
- Challenges: Robust place recognition is difficult because appearance changes with illumination, weather, season, time of day, viewpoint, shadows, and dynamic objects.These factors are especially critical when visually similar locations are only a few meters apart.
- Challenges: Visual information can complement GPS when measurements are noisy or unreliable around buildings, trees, narrow paths, and other sources of signal degradation.The introduction frames VPR as supporting location estimation rather than replacing GPS.
- YILDIZ-VPR contribution: YILDIZ-VPR provides dense, repeated walking traversals on the Davutpasa campus, covering historical and modern buildings, roads, green areas, open spaces, and wooded regions.The campus offers a compact but visually rich setting where natural and man-made structures coexist within a relatively dense spatial area.
2. YILDIZ-VPR Dataset
YILDIZ-VPR is a dense, pedestrian-level outdoor VPR dataset built from repeated campus traversals across varied times, seasons, weather, and illumination conditions. It provides synchronized GPS annotations, diverse train/test splits, and auxiliary sensor metadata for image-based, temporal, and sensor-aware localization research.
- Dataset design: YILDIZ-VPR uses repeated pedestrian-level traversals to capture dense outdoor visual coverage under changing times of day, weather, seasons, and illumination.The dataset targets realistic environmental variation rather than controlled single-session image matching.
- Collection environment: Collected on Yildiz Technical University’s Davutpasa campus, the dataset combines historical and modern buildings, roads, open and grassy areas, and wooded regions.Indoor areas were excluded to focus on outdoor place recognition, including challenging visually similar nearby locations.
- Geolocation and sampling: Approximately 1.5 meters average GPS accuracy, below-5-meter inclusion filtering, 8–9 Hz GPS recording, and approximately 4 Hz synchronized video sampling yield dense location annotations.The sampling strategy corresponds to roughly four images per meter depending on walking speed.
- Dataset organization: The dataset contains 370,464 training images, 2,377 daytime queries, and 1,902 nighttime queries organized for gallery-based retrieval and illumination-focused localization.Training frames are regularly sampled, whereas test queries are manually selected to reduce redundancy between adjacent frames.
- Sensor metadata and applications: Alongside RGB images and GPS coordinates, YILDIZ-VPR provides temperature, speed, and three-axis gyroscope metadata for motion and acquisition context.The video sequences and synchronized sensor data also support future temporal and sensor-aware place recognition studies.
3. Conclusion
YILDIZ-VPR is a densely sampled pedestrian-level dataset collected across varied temporal, seasonal, weather, and illumination conditions. Its dense coverage and long-term appearance variation support multiple visual place recognition settings, while future work will establish evaluation protocols and benchmark methods.
- Conclusion: YILDIZ-VPR contains 370,464 training images, 2,377 daytime query images, and 1,902 nighttime query images collected through repeated pedestrian-level traversals.The traversals covered different times of day, seasons, weather conditions, and illumination settings on the Davutpasa campus.
- Conclusion: YILDIZ-VPR combines dense spatial coverage, repeated observations, pedestrian-level viewpoints, and substantial long-term appearance variation.These properties provide complementary coverage for realistic visual place recognition research.
- Conclusion: The dataset supports image-based, sequential, day-to-night, and multimodal visual place recognition investigations.Its varied observations enable multiple recognition settings within the same resource.
- Conclusion: Future work will establish comprehensive evaluation protocols and benchmark representative visual place recognition methods on YILDIZ-VPR.The planned benchmarking is intended to evaluate representative VPR methods using the new dataset.