NAD: Neuron Activation based Divergence Maps for Weakly Supervised Object Localization

Abstract

Convolutional neutral networks (CNNs) have brought about massive improvements in the field of computer vision in solving some of the most conmplex problems like object detection, image captioning, semantic segmentation etc. These networks perform very well for such tasks but very little is known about why they do so. Their lack of transparency makes them difficult to interpret and that is why they are considered as black boxes. In this paper, we have proposed an approach in which we carry out weakly supervised object localization in images which eventually helps us understand the functioning of CNNs by providing us with the visual explanations for the predictions of CNNs. The proposed work focuses on exploiting the learned feature dependencies between consecutive layers of CNN. Different strategies are employed for different types of layers (Fully Connected layer, Convolutional layer etc.) to compute a binary value signifying neuron relevance. Moreover, we employ a method in which the computed activation maps corresponding to the non-target class are discounted from those of the target class in order to eliminate the irrelevant neurons and amplify the most discriminative neurons. This process highlights the most significant neurons of the CNN which have contributed the most in the prediction of a particular object. Our proposed approach performs better than the previously developed techniques with a better accuracy.