/VisualBackProp

VisualBackProp - visualization method for convolutional neural networks

Primary LanguageLua

VisualBackProp

VisualBackProp - visualization method for convolutional neural networks

Description

Detailed description of the VisualBackProp can be found in: https://arxiv.org/abs/1611.05418

Content:

  • inputImages - folder with input images
  • outputImages - folder for output images
  • vis.lua - implementation of VisualBackProp
  • run.lua - loads images, runs visualization method, saves images
  • model.t7b - model trained to predict steering wheel angle

Usage

mkdir outputImages

qlua run.lua

Script generates set of 3 images for each input image. Each set contain:

  • Input image with visualization mask overlaid in red
  • Averaged feature maps
  • Intermediate masks