R-FCN: Object Detection via Region-based Fully Convolutional Networks

R-FCN is a region-based, fully convolutional network for object detection. Unlike Fast and Faster R-CNN, which apply an expensive per-region subnetwork hundreds of times, R-FCN shares almost all computation across the entire image. To reconcile the tension between the translation-invariance needed for image classification and the translation-variance needed for detection, the authors introduce position-sensitive score maps, which allow the detector to adopt fully convolutional image-classification backbones such as Residual Networks.

On the PASCAL VOC datasets, R-FCN with a 101-layer ResNet achieves competitive accuracy, reaching 83.6% mAP on the 2007 set, while running at 170ms per image at test time, 2.5 to 20 times faster than the Faster R-CNN counterpart. By combining strong accuracy with substantially higher speed and releasing code publicly, the work showed that fully convolutional designs could deliver efficient, accurate detection.