Single-Shot Cascade Bounding Box Refinement Neural Network for High Quality Object Detection

This paper proposes a novel one-stage detection method with cascade bounding box refinement, called CasDet. Generally one-stage detection methods generate less accurate bounding boxes than two-stage methods. Motivated by the location refinement of bounding boxes of two-stage methods, we present a cascade bounding box refinement for one-stage detection networks. By refining the detection results for multiple times, CasDet can generate high quality bounding boxes. To further improve the feature representation, we present a multi-scale feature aggregation structure (MBFA). By fusing features with adjacent three scales and adding multiple top-down and bottom-up aggregation paths, MBFA introduces rich contextual information and local details into each detection feature. The experimental results on COCO benchmark confirm that our CasDet outperforms existing state-of-the-art detection methods and achieves a better balance between speed and accuracy.