[Paper] Backprop: Visualising Image Classification Models and Saliency Maps (Weakly Supervised Object Localization)

Weakly Supervised Object Localization (WSOL) Using AlexNet

Visual Geometry Group, University of Oxford

Outline

1. Gradient-Based Class Model Visualisation

Gradient-Based Visualization (different aspects of class appearance are captured in a single image)

2. Image-Specific Class Saliency Visualisation

Image-specific class saliency maps (The maps were extracted using a single back-propagation pass through a classification ConvNet. No additional annotation.)

3. Weakly Supervised Object Localization (WSOL)

3.1. Segmentation Using GraphCut

GraphCut colour segmentation
Left: Image, Left-Middle: Saliency Map, Right-Middle: thresholded saliency maps blue shows the areas used to compute the foreground colour model, cyan — background colour model, pixels shown in red are not used for colour model estimation. Right: Segmentation results

3.2. ILSVRC-2013 Localisation Challenge

PhD, Researcher. I share what I've learnt and done. :) My LinkedIn: https://www.linkedin.com/in/sh-tsang/, My Paper Reading List: https://bit.ly/33TDhxG