Hierarchical class grouping with orthogonal constraint for class activation map generation

Class activation map (CAM) generation aims at highlighting regions of a class in an image by the classification model. However, the regions obtained are usually small and local. Existing methods attribute the problem to the ineffective CAM extraction model and pay much attention on enlarging the regions via developing new models for CAM generation, but limited success has been made. Different from the existing methods, this paper attributes such incompleteness extraction to the finite discriminative cues within a single classification model and improves CAM generation by providing more discriminative cues via training multiple classification models with consideration of class relationships. To this end, the similarities between classes are firstly measured, and hierarchical clustering is then implemented to cluster initial clusters into multiple semantic meanings level of clusters. Afterward, multiple classification models are trained on these different levels of clustering, and multiple class activation maps with various and complementary discriminative cues are obtained. Finally, the class activation map is obtained via the combination of these maps. A new orthogonal module and a two-branch network for CAM generating are also proposed to improve CAM generation via making the regions orthogonal and complementary. Experimental results on the PASCAL VOC 2012 dataset show the superior performance of the proposed CAM generation method.

Automated summary

This paper proposes an improved CAM generation method by training multiple classification models and using hierarchical clustering to obtain more discriminative class activation maps. Experimental results show superior performance of the proposed method on the PASCAL VOC 2012 dataset.