GTA: Guided Transfer of Spatial Attention from Object-Centric Representations

Utilizing well-trained representations in transfer learning often results in superior performance and faster convergence compared to training from scratch. However, even if such good representations are transferred, a model can easily overfit the limited training dataset and…