Rank and Sort Loss-Aware Label Assignment with Centroid Prior for Dense Object Detection

The recent progress in object detection seeks to design more effective and dynamic label assignment strategies that automatically select training samples in a prediction-aware manner. In this paper, we revisit the loss-aware label assignment and innovatively propose the Rank & Sort (RS) Loss-aware Label Assignment with Centroid Prior (RSLLACP), which is more noise-robust and adapted to the semantic patterns of each instance. By taking advantage of the instance mask annotation, the centroid prior is more appropriate than the geometric center to define the region for positive anchors due to more informative features contained within. Besides, the centroid prior prevents the ambiguous anchors from taking place. Inspired by the recent advances that the ranking-based objective functions can dramatically improve the detection performance, RSLLACP proposes to incorporate the RS cost into the matching cost matrix to replace the classification cost. Thanks to its rankingbased nature, the positive anchors are differentiated from the negatives by the classification logits while being robust to the foreground-background class imbalance. Due to its sorting objective, positive anchors are prioritized with respect to their continuous localization qualities. This ranking and sorting nature lines up with the label assignment objective. Extensive experiments on the MS COCO dataset validate the effectiveness of our proposed RSLLACP. Without bells and whistles, RSLLACP achieves 51.9 AP, outperforming all existing state-of-the-art one-stage detectors by a significant margin.