Clustering Anchor for Faster R-CNN to Improve Detection Results

Object detection has made impressive progress in recent years where Faster R-CNN is the mainstream framework for region-based object detection methods. However, a single Faster R-CNN framework no longer has advantages compared with the latest detection models. So based on Faster R-CNN, a model that focuses on features, normalization methods, and anchor sizes is proposed to improve detection results. The model integrates Feature Pyramid Networks (FPN), Group Normalization (GN) with k-means clustering. FPN is used to produce a multi-scale feature representation, which enables the model to detect objects across a wide range of scales. GN addresses the problem of the small training batch size effectively. K-means clustering algorithm is used finally to determine anchor sizes of the network on the purpose of making the network do bounding-box regression more easily. Without bells and whistles, the detection model achieves state-of-the-art object detection accuracy on the MSCOCO datasets.