Entropy regularization for weakly supervised object localization

Abstract

The goal of weakly-supervised object localization (WSOL) is to train a localization model without the location information of the object(s). Recently, most existing WSOL methods capture the object with an attention map extracted from a classification network. However, it has been observed that we need to sacrifice classification performances to achieve the best WSOL score. We conjecture that this is because the objective of classification training, minimizing entropy between one-hot ground truth and predicted class probability, is not entirely consistent with that of localization. In this paper, we investigate how the entropy of predicted class probability affects localization performances, where we conclude that there is a sweet spot for localization with respect to entropy. Hence, we propose a new training strategy that adopts entropy regularization for finding the optimal point effectively. Specifically, we add an additional term to the loss function, which minimizes the entropy between a uniform distribution and the predicted class probability vector. The proposed method is easy to implement since we do not need to modify the architecture or data pipeline. In addition, our method is efficient in that almost zero additional resources are required. Most importantly, our method improves WSOL scores significantly, which has been shown through extensive experiments.