Tweak image featurization to use "prob" layer and max-probability dot product matching logic.
IPython notebook tests using the night/day:concert/riot data sets found improved results with some minor changes to our image featurization logic. Changes needed:
Currently use "fc7" layer of the cafe model, switch to "prob"
Currently aggregate overall cluster vector via averaging each sub-element, switch to selecting max value of each element.
Currently compare aggregate cluster vector to candidate using cosine similarity, switch to using dot-product.
Tweak image featurization to use "prob" layer and max-probability dot product matching logic.
IPython notebook tests using the night/day:concert/riot data sets found improved results with some minor changes to our image featurization logic. Changes needed: