Learning Deep Features for Discriminative Localization

TL;DR

This paper demonstrates that global average pooling in CNNs enables effective localization of objects using only image-level labels, achieving near-supervised performance on localization tasks.

Contribution

It reveals that global average pooling creates a versatile localizable deep representation, significantly improving weakly supervised localization capabilities.

Findings

Achieved 37.1% top-5 error for object localization on ILSVRC 2014

Global average pooling enables CNNs to localize discriminative regions without explicit localization training

Network performs well on various localization tasks despite being trained only on image labels.

Abstract

In this work, we revisit the global average pooling layer proposed in [13], and shed light on how it explicitly enables the convolutional neural network to have remarkable localization ability despite being trained on image-level labels. While this technique was previously proposed as a means for regularizing training, we find that it actually builds a generic localizable deep representation that can be applied to a variety of tasks. Despite the apparent simplicity of global average pooling, we are able to achieve 37.1% top-5 error for object localization on ILSVRC 2014, which is remarkably close to the 34.2% top-5 error achieved by a fully supervised CNN approach. We demonstrate that our network is able to localize the discriminative image regions on a variety of tasks despite not being trained for them