[Paper Review] Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps

Karen Simonyan, Andrea Vedaldi+1 more|arXiv (Cornell University)|Dec 20, 2013

Visual Attention and Saliency Detection10 references4,905 citations

TL;DR

The paper presents gradient-based visualization methods for CNNs: (1) generating class-representative images by optimizing input, and (2) producing image-specific saliency maps for weakly supervised localization, relating these to DeconvNet reconstructions.

ABSTRACT

This paper addresses the visualisation of image classification models, learnt using deep Convolutional Networks (ConvNets). We consider two visualisation techniques, based on computing the gradient of the class score with respect to the input image. The first one generates an image, which maximises the class score [Erhan et al., 2009], thus visualising the notion of the class, captured by a ConvNet. The second technique computes a class saliency map, specific to a given image and class. We show that such maps can be employed for weakly supervised object segmentation using classification ConvNets. Finally, we establish the connection between the gradient-based ConvNet visualisation methods and deconvolutional networks [Zeiler et al., 2013].

Motivation & Objective

• Understand what ConvNets learn about visual classes by visualizing class models as representative images.
• Develop image-specific saliency maps to identify regions contributing to class scores in a given image.
• Demonstrate that saliency maps enable weakly supervised object localization without extra annotations.
• Establish theoretical connections between gradient-based visualization and DeconvNet reconstructions.

Proposed method

• Compute class scores S_c(I) and optimise the input image I to maximise S_c(I) under L2 regularization.
• Back-propagate to obtain first-order derivatives w = dS_c/dI at a given image to form the image-specific saliency map M with M_ij = |w_h(i,j)| (grey) or M_ij = max_c |w_h(i,j,c)| (color).
• Use a single back-propagation pass to generate saliency maps for a chosen class in an input image.
• Apply GraphCut with foreground/background color models derived from saliency thresholds to achieve weakly supervised object localization.
• Relate gradient-based visualization to DeconvNet by showing approximate reconstructions correspond to gradient back-propagation through the network.

Experimental results

Research questions

• RQ1 Can gradient-based optimisation of the input reveal a CNN's notion of a class?
• RQ2 Can image-specific saliency maps uncover spatial support of a class in a given image without extra annotations?
• RQ3 Are gradient-based visualisation methods equivalent or related to DeconvNet reconstructions?
• RQ4 Can saliency-driven localization reach competitive performance under weak supervision?
• RQ5 How do visualisations inform understanding and potential integration into learning frameworks?

Key findings

• Class-model visualisation yields images that represent the visual appearance learned for a class by maximising the class score S_c(I).
• Image-specific saliency maps created from input derivatives highlight discriminative regions for a given class in an image and can be produced by a single back-propagation pass.
• Saliency maps enable weakly supervised object localization by thresholding and using GraphCut to produce object masks, achieving 46.4% top-5 localisation error on ILSVRC-2013 in their setup.
• The gradient-based visualisation generalises the DeconvNet reconstruction procedure, showing the equivalence or close relation between back-propagated gradients and DeconvNet reconstructions for most layers.
• The network used achieves 39.7% top-1 and 17.7% top-5 error on ILSVRC-2013 validation, improving over a referenced prior single-CNN result (40.7%/18.2%).