Cascade Feature Pyramid Neural Nwtwork for Object Detector

Early object detection networks apply single scale feature to predict the classification and regression. However, the semantic information which is from the low-level features is less than that on the high-level features, and locations of objects on the low-level are more accurate than that on the high-level. Recently, feature pyramid fusion is widely applied by one-stage object detectors (e.g. DSSD, RefineDet), which achieves encouraging results of the object detection, but only simply constructs the feature pyramid according to multi-scale feature maps. In our work, we design a network, called CFPNet, which applies cascade feature pyramid layer (CFP) to predict the classification and regression. CFPNet achieves 80.1% mAP with small input size (320× 320) and 82.0% mAP with large input size (512 × 512), experimenting on PASCAL VOC2007 and PASCAL VOC2012. The network with low dimension input size runs 34 FPS in NVIDIA 1080Ti, which achieves the high accuracy and maintains fast speed.