A Weakly Supervised Object Localization Method Based on Region-weight Smoothing

Weakly supervised object localization methods with category supervision suffer from the problem that they tend to merely cover the most discriminative components of the object. And the category response of the region is affected by key weights, and the imbalance of them leads to the sparsity of object location. So this paper proposed a solution based on region-weight smoothing. This paper designed an adaptive standard deviation regularization to shrink weights discrepancy, which could smooth object location while preserving the classification performance. Results of experiments on several datasets show that this method could generate a wider area and achieve higher accuracy.