Forget And Diversify: Regularized Refinement For Weakly Supervised Object Detection

We study weakly supervised learning for object detectors, where training images have image-level class labels only. This problem is often addressed by multiple instance learning, where pseudo-labels of proposals are constructed from image-level weak labels and detectors are learned from the potentially noisy labels. Since existing methods train models in a discriminative manner, they typically suffer from collapsing into salient parts and also fail in localizing multiple instances within an image. To alleviate such limitations, we propose simple yet effective regularization techniques, weight reinitialization and labeling perturbations, which prevent overfitting to noisy labels by forgetting biased weights. We also introduce a graph-based mode-seeking technique that identifies multiple object instances in a principled way. The combination of the two proposed techniques reduces overfitting observed frequently in weakly supervised setting, and greatly improves object localization performance in standard benchmarks.