In this work, a first step towards the direct detection of “good” features using a fully convolutional neural network (CNN) is presented. By designing a convolutional light field autoencoder and comparing the results from 2D and 3D convolutions, we show that
convolutions can pick up on additional information in the light field in higher dimensions.
We then build a CNN to detect SIFT features on 2D input images. We show that the model architecture can be adapted to work with higher-dimensional inputs.
In this work, we present a learning approach for feature detection. We show that our method, using a fully convolutional neural network, is feasible for two-dimensional images and generalizes to the higher dimensional light fields.