Learning Deep Features for Discriminative Localization

Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, Antonio Torralba
Computer Science and Artificial Intelligence Laboratory
MIT

回目錄

Abstract

1. Revisit the GAP
   • GAP(global average pooling layer)
2. Achieve 37.1% top-5 error on ILSVRC 2014
   • ILSVRC(Large Scale Visual Recognition Challenge)

Introduction

1. convolutional layers can localize objects, but this ability is lost when fully-connected layers
   • Network in Network, GoogLeNet avoid fully-connected layers.
     • minimize the number of params
2. GAP(known as a kind of structural regularizer) doesn't simply act as a regularizer.
3. This approach can be easily transferred to other recognition datasets for generic classification, localization and concept discovery.
   • achieves 37.1% top-5 test error, close to the fully supervised AlexNet.

Related Work

Class Activation Mapping

Weakly-supervised Object Localization

Setup

• use AlexNet, VGGnet and GoogLeNet to generate *GAP
  1. remove some layer (fully connected layer and softmax)
  2. add some convolution layer
     • 3 * 3, stride 1, padding 1 with 1024 units.
     • GAP layer
     • softmax

Results

1. Classification
   

• similar

2. Localization
   

• weakly supervision still have a long way to go.

3. Generic Localization

• to test the ability about feature extraction between original network and the network concatenated with GAP by linear SVM
  • similar
• to test the ability about localization by weakly supervision
  • It still can find the position of the object.

Deep Features for Generic Localization

Fine-grained Recognition

• with bounding box

Pattern Discovery

• Given a set of images containing a common concept, test the network whether can find where the position f the important regions in this images.
  1. How to identify the important region before train the network to test the network performance.
     1. use GoogLeNet-GAP network training by image-level label. use SVM weight and GAP to contruct the CAM to identify the important region.
  2. Experiment

     1. Discovering informative objects in the scenes

     2. Concept localization in weakly labeled images

     3. Weakly supervised text detector

        • postive set: picture with text
        • negtive set: picture without text

     4. Interpreting visual question answering (???)

Visualizing Class-Specific Units

Conclusion

1. CAM fofr CNN with global average pooling.
   • enable to visualize hotmap to the given image
2. weakly supervision to find localize the object.