Download
R-FCN (“Region-based Fully Convolutional Networks”) is an object detection framework that makes almost all computation fully convolutional and shared across the image, unlike prior region-based approaches (e.g. Faster R-CNN) which run per-region sub-networks. The repository provides an implementation (in Python) supporting end-to-end training and inference of R-FCN models on standard datasets. The authors propose position-sensitive score maps to reconcile the need for translation variance (in detection) and translation invariance (in classification). R-FCN is efficient (low per-region overhead) and competitive in accuracy (e.g. with ResNet backbones). Position-sensitive score maps for per-region classification without expensive per-region convs. Optional “deformable R-FCN” extension for improved performance.
Features
- Fully convolutional design with shared feature extraction across the image
- Position-sensitive score maps for per-region classification without expensive per-region convs
- End-to-end trainable pipeline (proposal + classification)
- Support for multiple backbone architectures (e.g. ResNet)
- Optional “deformable R-FCN” extension for improved performance
- Low per-RoI overhead (fast inference)
Project Samples
