A Novel Multi-scale Key-Point Detector Using Residual Dense Block and Coordinate Attention

Object detection, one of the core missions in computer vision, plays a significant role in various real-life scenarios. To address the limitations of pre-defined anchor boxes in object detection, a novel multi-scale key-point detector is proposed to achieve rapid detection of natural scenes with high accuracy. Compared with the method based on key-point detection, our proposed method has fewer detection points which are the sum of pixels on four-layer compared to one-layer. Furthermore, we use feature pyramids to avoid ambiguous samples. Besides, in order to generate feature maps with high quality, a novel residual dense block with coordinate attention is proposed. In addition to reducing gradient explosion and gradient disappearance, it can reduce the number of parameters by 5.3 times compared to the original feature pyramid network. Moreover, a non-key-point suppression branch is proposed to restrain the score of bounding boxes far away from the center of the target. We conduct numerous experiments to comprehensively verify the real-time, effectiveness, and robustness of our proposed algorithm. The proposed method with ResNet-18 and resolution of \(384\times 384\) achieves \(77.3\%\) mean average precision at a speed of 87 FPS on the VOC2007 test, better than CenterNet under the same settings.