CenterNet Based On Gradient Harmonizing Mechanism

Anchor free methods for object detection such as CornerNet and CenterNet proposed in the past two years,are lightweight models and have the characteristic of one to one correspondence between positive samples and ground truths.However,due to the imbalance of the easy-splitting and hard-splitting samples as well as the imbalance of positive and negative samples and because it's hard to correspond each output of head networks with others,the average precisions of anchor free methods are not high enough.In this paper.We improve three head networks of CenterNet,GHM can reduce the impact of easy-splitting and hard-splitting samples on loss function.We propose two CenterNet models with GHM-R loss and GHM-C loss.Numerical experimental results show that these two new models have higher average precision than CenterNet.In addition,we improve focal loss and gradient harmonizing mechanism.We set the modulating factor of the positive samples in the focal loss to 1,which improves the contribution of the positive sample to the loss.We defined half focal loss,and improve the model.And the ?-gradient density is changed to the m-gradient density,which solves the problem that the gradient density harmonizing parameters may be unbounded.We also add upper and lower bounds that vary with epoch to the gradient density harmonizing parameters,witch named bounded gradient harmonizing mechanism,so that the loss function smoothly transitions from focal loss to gradient harmonizing mechanism.Numerical experimental results show that CenterNet with BGHM has a higher average precision.