This work uses a subset of the Global Streetscapes dataset. The original SVIs are downloaded following their wiki. Their methodology can be found in the paper.

First, the same points as the Global Streetscapes dataset from the chosen five cities, Singapore, Amsterdam, San Francisco, Abuja, and Santiago, are chosen

Use the notebook 1_contextual-filtering.ipynb to filter the selected points based on contextual attributes such as panorama status, lighting condition, and season. The notebook 1b_visual-complexity-analysis.ipynb sets thresholds of visual complexity for the different cities.

Use the notebook 2_segmentation-clustering.ipynb to first calculate the visual complexity of each image based on the segmentation results. Images with low visual complexity are suspected to have one predominant feature such as a wall or highway. Thus, these images are not considered.

Then, compute the pixel ratio of diverse elements (i.e., road, vegetation, car, building, and sky) and cluster the SVIs based on them. The pixel ratios are calculated for each city and the optimal number of clusters, using this five features, and K-means is found by trying k = [2, 11] and choosing k with the highest average Silhouette score.

Use the notebook 3_random-sample.ipynb to randomly sample SVIs for each city and from each cluster. The number of SVIs per city on this work are 80. The final dataframe, made out of 80 randomly sampled SVIs per city, should be saved in download_imgs/filtered_points.csv.

Finally, move into the folder download_imgs/ and download the respective SVIs of the filtered points.

python download_imgs.py

The images were manually reviewed, and the notebooks 4_resample_mising_<CITY>.ipynb, where <CITY> is the city name, were run until 80 images were chosen.

Finally, the notebook 5_generate_final_svis_metadata.ipynb is used to have a consolidated metadata csv of the final images.