Anchor-free object detection network based on non-local operation.

Keypoint detector has achieved satisfactory performance in facial recognition and computer vision-related fields. However, convolutional neural network are somewhat limited for capturing global information, which greatly affects the performance of the detector. This paper proposes an efficient solution which obtains additional global information at a minimal costs. The framework is built upon a representative one-stage keypoint-based detector named CenterNet. The global keypoint network(GKPNet) approach forms a larger receptive field by using non-local modules to obtain more information than stacking multiple convolution modules. Feature pyramid network(FPN) was improved to extract multi-scale information and fuse low-level detail information to enhance the detection of small objects. The proposed approach is efficient and achieve 80.3% average precision on the Pascal VOC validation dataset, thus outperforms the CenterNet detector at least 1.6%. With a higher inference speed, GKPNet demonstrates performance comparable to that of the two-stage detectors and thereby provides a further avenue for anchor-free object detection. [ABSTRACT FROM AUTHOR]