TL;DR
This paper introduces the Virtual City Explorer (VCE), a system that leverages paid crowdworkers and street-view imagery to map Points of Interest (PoIs) in urban areas. They demonstrate that this approach is a scalable and cost-effective alternative to traditional methods, even identifying PoIs missed by experts, and introduce a "taboo" mechanism to improve exploration efficiency.
What Problem Does This Paper Address?
Mapping urban environments and identifying Points of Interest (PoIs) is crucial for various applications, from urban planning to navigation. Traditional methods like expert surveys are expensive and time-consuming, while Volunteered Geographic Information (VGI) relies on the availability and geographic proximity of volunteers, limiting scalability. This paper addresses the challenge of efficiently and accurately mapping PoIs over large areas by exploring the use of paid crowdworkers who can remotely access and analyze street-view imagery, overcoming the limitations of volunteer availability and location.
Methods
The core of the paper is the Virtual City Explorer (VCE) system. It comprises three main components:
- Configurer: Allows end-users to define the area of interest, the type of PoI to locate, and the number of crowdworkers.
- Exploration Task Engine: Presents workers with street-view imagery and instructions, guiding them to explore a defined area. Workers report PoI locations by taking three photos from different angles, which are then validated for triangulation.
- Aggregator: Consolidates reported PoI coordinates using the DBSCAN algorithm to cluster nearby detections and form a final map.
A key innovation is the "Taboo" optimization. Once a PoI is independently confirmed by a certain number of workers, it becomes "taboo" for subsequent workers, encouraging them to explore new areas. An "escape condition" is also implemented to reward workers who spend significant time and effort without finding PoIs, preventing unfairness.
Datasets
The VCE was evaluated using two real-world areas:
- Trento - ZTL (Zona a Traffico Limitato), Italy: A 0.347 km² area with a pre-existing bike rack map created by a municipal employee.
- Penn Quarter, Washington D.C., USA: A 0.33 km² area with bike rack data from the VGI mobile app "Rackspotter."
The authors also compared their results against OpenStreetMap (OSM) data for both areas.
Key Insights and Ideas
- Street-view imagery + paid crowdworkers = scalable PoI mapping: This combination effectively overcomes the geographical limitations of VGI and the cost/time constraints of expert surveys.
- Complementarity with existing data: The VCE can identify PoIs missed by expert surveys or VGI communities, suggesting it's a valuable complementary tool.
- The "Taboo" strategy boosts exploration: Making confirmed PoIs taboo significantly encourages workers to explore new areas, leading to more discoveries and potentially better coverage.
- Worker behavior is key: Understanding how workers interact with the interface, their error patterns, and their exploration strategies is crucial for improving the system.
Key Results
- Feasibility and Completeness (RQ1): Paid crowdworkers were effective in identifying PoIs, with a low rate of false positives due to the aggregation component. While not always finding all PoIs, they discovered PoIs missed by other methods.
- Area Coverage (RQ2): The VCE achieved high area coverage (around 90% for Trento, 97% for Washington) with 60 task executions.
- Cost and Time (RQ3): The VCE approach was found to be significantly cheaper than on-field surveys and faster than OSM for generating maps of specific areas, though more expensive than free VGI alternatives.
- Taboo Optimization Impact (RQ4): The Taboo strategy increased confirmed PoI detection by 30% in Trento and less than 10% in Washington, while reducing the number of redundant detections.
- Worker Behavior (RQ5): Most workers completed tasks without significant errors. The Taboo strategy did not significantly increase task completion time but did lead to increased walked distance in Trento, suggesting more thorough exploration.
Limitations and Future Work
- Completeness is not absolute: The VCE doesn't find all PoIs, and the number of detected PoIs depends on task executions and worker effort. Future work could involve dynamic adjustment of these parameters.
- Street-view imagery limitations: The approach is constrained by the availability and recency of street-view imagery.
- Dynamic starting points: The authors suggest exploring dynamic starting point functions to guide workers into unexplored areas.
- Alternative exploration strategies: Comparing the current free-exploration approach with guided segment exploration is a future direction.
- Dynamical reward strategies: Investigating reward mechanisms that adapt to the number of PoIs reported could further optimize the system.
- Quality assessment without ground truth: Developing methods to assess agreement among workers for quality assurance in the absence of reference datasets is a key future step.
Summary updated