Enhanced Task-Aware Spatial Disentanglement Head for Oil Tanks Detection in High-Resolution Optical Imagery

Abstract

Task-aware spatial disentanglement (TSD) head could decouple classification and regression from the spatial dimension by using deformable pooling. However, the effect of deformable pooling decreases when the Intersection over Union (IoU) of the positive sample increases as the deformable pooling destroys the learned Region of Interest’s (RoI’s) pooling feature. Besides, deformable pooling is to find the useful area of an object to predict the object in a proposal, so deformable pooling is not suitable to some extent for negative samples as deformable pooling will find the useful area of an object to predict there is no object in negative samples. We propose an enhanced TSD (ETSD), which adds an auxiliary branch with IoU-based loss weight. The auxiliary branch has double identical structure heads to separate parameters for classification and regression and uses the RoI’s pooling features as input to train classification and prediction directly. The loss weight of the auxiliary branch will be doubled for negative samples and increased gradually for positive samples according to the IoU. Double heads in the auxiliary branch make the RoI’s pooling feature get optimal feedback in backpropagation. The increase in the loss weights of the auxiliary branch will allow the model to enhance the training of the RoI’s pooling feature. The more accurate the RoI’s pooling feature, the more accurate the deformable pooling features learned from the RoI’s pooling feature. We performed detailed ablation studies on the oil tank dataset extracted from DIOR.