Traffic signs are a key feature for navigating and managing traffic safely, affecting all of us on a daily basis. However, traffic sign datasets are lacking on open government data portals as well as OpenStreetMap (OSM). Mapillary’s computer vision capabilities can extract more than 1,500 classes of traffic signs globally from street-level imagery. Generated traffic signs are available on iD Editor, Rapid and JOSM Mapillary plugin to enrich OpenStreetMap data. Our team wanted to know how the accuracy of traffic signs detected by Mapillary compared with the reality on the ground (the ground truth). To answer this question we collected more than thousands ground truth data in San Francisco and used this information to produce the recall, precision, and positional accuracy of our machined generated traffic sign data. This provided some interesting insights in OpenStreetMap and the level of completeness and gaps of that dataset. In this talk, we will cover Mapillary’s traffic sign extraction capabilities, Mapillary generated traffic sign data against ground truth data and OSM’s traffic sign coverage in San Francisco’s downtown. We will be also addressing how data quality can be improved using various data collection techniques and the role of post-processing with Structure from Motion and control points annotations.
None
Speakers: Yunzhi Lin Said Turksever