Modelpublic

Spatial-Data-Science-and-GEO-AI-Lab/svi_percept

sourceHugging Facegpl-3.0updated 2y agoView on Hugging Face

0likes

Model Card

K-Nearest-Neighbour model of human perception of Amsterdam street view imagery. data.npz is a compressed numpy file with 10 records, it can be loaded with numpy.load and the resulting object can be indexed by any of the following keys:

—walkabilityvecs, walkabilityscores
—bikeabilityvecs, bikeabilityscores
—pleasantnessvecs, pleasantnessscores
—greennessvecs, greennessscores
—safetyvecs, safetyscores

The vecs entries are matrices of size Nx1024 and the scores entries are vectors of size N.

The _vecs are encoded by OpenCLIP ViT-H-14-378-quickgelu (pretrained: dfn5b).

The scores are the given rating scores (1 to 5) for the corresponding vector. For example, the vector given by walkabilityvecs[10,:] has a corresponding score in walkability_scores[10].

This can be used to model a score for any given vector from an image encoded by the above CLIP model.

Example spaces

—Percept-map: pick a point on a map and evaluate perception ratings for Mapillary imagery from that location
—Percept-image: evaluation perception ratings for a given image