Similar to the approach for Pinterest visual search, but compares the feature outputs of a CNN directly without creating a 'visual vocabulary'. The search strategy is k approximate nearest neighbors, with k as a search parameter, allowing for a speed/accuracy tradeoff. Currently, I'm exploring offering this as a service, and am working on customer/product development and open to any and all input and suggestions.
1 comment
[ 2.9 ms ] story [ 9.5 ms ] thread