The paper presents a better loss function for anchor-based detection methods by matching anchors to GT boxes in a differentiable manner. Results are convincing. Three reviewers recommend acceptance after a convincing rebuttal. The final decision is to accept.