Reverse Image Search (Fast)

This app demonstrates how to quickly retrieve the k-nearest neighbors to an image query. It is the same concept as what’s given in the Reverse Image Search, except that it can make queries on previously-unseen images at nearly real-time, whereas the former is only a viewer for a fixed set of images pre-analyzed by a Python script. It is also mostly identical to Reverse Object Search Fast, except that it operates over the entire image rather than over detected objects within the image. Live reverse image search at ~1 fps A large dataset of images is pre-analyzed by a convolutional neural network trained to classify images. A feature vector is extracted from each image, the 4096-bit activations from the last fully-connected layer (just before the classification layer), and recorded, along with the image’s original file path. After all feature vectors are extracted, they are transformed into a lower dimensional representation (for dimensionality reduction) using a random matrix project...

1 mentions: @genekogan
Date: 2019/06/12 06:47

Referring Tweets

@genekogan @hardmaru a nice example of the latter: when i was working on I found out that you could get similar quality by replacing PCA with random linear projection of the feature vectors