Label Assignment for Aligning Features in Anchor-free Models

Most anchor-free methods perform object detection in the way of dense recommendation, which assumes one point can conduct high-quality category prediction and regression estimation simultaneously. However, due to different task drivers, valid classification and regression features may not locate around the same sampling point in the training phrase. This problem is called feature misalignment. To solve it, this paper proposes a new label assignment strategy with feature filter to guide the network to focus on sampling points with aligned features. In addition, we establish mutually independent multi-layer quality distributions to model objects' priori information on different feature pyramid scale levels. Equipped with our method, the generated foreground weight map gathers more to centers of the classification and regression heatmaps. Experiment results show that without bells and whistles, our method achieves 45.3% AP on COCO test-dev under the default 2x schedule, outperforming other related methods.