TextureToMTF: Predicting Spatial Frequency Response In The Wild

The Modulation Transfer Function(MTF) is often used to quantify the spatial resolution of a camera. Typically, we use the value known as MTF50, which is the minimum spatial frequency at which the magnitude of the frequency responseis half its value at zero frequency (i.e. the DC component). An efficient way of computing the MTF50 value is by using the Spatial Frequency Response(SFR). Traditionally, SFR has been captured using specialized charts and controlled setups. These setups are specifically designed to make the calculations invariant to rotation, scale, illumination etc. The use of these charts, however, is limited by the fact that laboratory setups do not always correlate well with real world environment, due to reasons like uncontrolled illumination, varying depth in the scene etc. In this poject, we investigated the problem of predicting spatial frequency response ‘In thewild’, i.e. for casually captured images in uncontrolled real world scenarios.

code  /  dataset  /  pdf

Local Blur Detection

  • Images shows local MTF50 values in the form ofa heat map. The top row shows examples of natural image with varying texture. The bottom row shows the heat maps generated for the corresponding images using our network. The warmer colors represent higher MTF50 scores(better quality).