A pedestrian perspective street view dataset for prosthetic vision simulation.

Traditional street view datasets are collected from a vehicle-mounted perspective predominantly, which cannot truly reflect pedestrians' perception in complex urban environments. A Pedestrian Perspective Street View Dataset (PPSVD) for the study of simulated prosthetic vision was proposed and constructed in this paper to maximize the reproduction of street scenes from a pedestrian's perspective and meet the complex environmental simulation needs for simulated prosthetic vision research. There were 1200 high-quality instance-level annotations' images collected from real street scenes to form PPSVD which covered 16 typical object categories of common street elements such as pedestrians, vehicles, buildings or traffic lights etc. The experimental results showed for the segmentation accuracy of target areas, the model was trained on the PPSVD significantly outperformed than those trained on general datasets. Especially, in complex scenes, small targets and occluded situations, the performance of the model was improved even more pronounced. Furthermore, the PPSVD demonstrated significant performance improvements across models of diverse sizes and structures, validating its effectiveness and applicability in various algorithms. Therefore, the proposed PPSVD not only creates a more realistic environment for simulated prosthetic vision experiments but also serves as a crucial data resource for related research.