SIFT Features for Deep and Variational Autoencoders: A Performance Comparison

Object detection tasks widely use the Scale Invariant Feature Transform (SIFT). SIFT key-points are extracted from model images of an object and are used to form descriptors. To achieve an object's existence from a query image, we find the candidate's matching features based on their feature vector's Euclidean distance. The goal is to maximize the SIFT features so that classification can be more accurate and reliable. In this work, we compare the encoding efficiency using the number of SIFT matches found in the images of Deep Autoencoder with that of Variational Autoencoder.