Faster R-CNN
Faster Region-based Convolutional Neural Network · Also known as: Faster RCNN, Faster-RCNN, RPN-based detector, two-stage object detector, region proposal network detector
Faster R-CNN is a two-stage deep convolutional object detection framework introduced by Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun (Microsoft Research) at NeurIPS 2015. It replaces the slow selective-search region proposal step used in its predecessors R-CNN and Fast R-CNN with a learned Region Proposal Network (RPN) that shares convolutional features with the detection head, enabling the first end-to-end trainable, near-real-time accurate object detector and establishing a long-standing accuracy benchmark on PASCAL VOC and MS COCO.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Faster R-CNN is the right choice when detection accuracy is the primary requirement and moderate inference latency (tens of milliseconds per image on a GPU) is acceptable — for example in medical imaging, satellite image analysis, industrial inspection, or dataset annotation pipelines. It is well-suited to detecting small or densely packed objects where the RPN's multi-scale anchors provide coverage that single-stage detectors may miss. A GPU is effectively required: inference on CPU is prohibitively slow for real-world use. For applications requiring real-time throughput at 30+ FPS on a single GPU, single-stage detectors such as YOLO or SSD are more appropriate. A labelled dataset with bounding-box annotations is mandatory; the model is typically initialised from an ImageNet-pretrained backbone and fine-tuned, so even a few thousand annotated images can be sufficient with transfer learning.
Strengths & limitations
- State-of-the-art detection accuracy (mAP) at its introduction, and still competitive with proper backbones.
- End-to-end trainable: the RPN and detection head share features and are optimised jointly, eliminating the ad-hoc proposal pipeline of earlier detectors.
- Multi-scale anchors enable robust detection of objects across a wide range of sizes and aspect ratios.
- Modular backbone: any ImageNet-pretrained CNN (VGG, ResNet, etc.) can be plugged in as the feature extractor.
- Served as the foundation for many subsequent architectures including Mask R-CNN, Feature Pyramid Networks, and Cascade R-CNN.
- Inference is slower than single-stage detectors (YOLO, SSD, RetinaNet) at comparable accuracy, limiting use in strict real-time applications.
- The two-stage pipeline adds implementation complexity and more hyperparameters (anchor sizes, NMS thresholds, RPN sampling ratios) compared to single-stage methods.
- RoI Pooling introduces quantisation artefacts for small regions; RoI Align (introduced in Mask R-CNN) corrects this but requires replacing the pooling layer.
- Memory and compute requirements are substantial; training from scratch without a pretrained backbone is impractical for most practitioners.
- Performance degrades on very densely overlapping objects where NMS may suppress true detections.
Frequently asked
What are anchors and why does Faster R-CNN use them?
Anchors are a set of fixed reference bounding boxes of different scales and aspect ratios tiled over every spatial location of the feature map. Rather than searching over all possible box positions and sizes, the RPN only needs to predict small offsets from these reference boxes, which makes the regression problem much easier and ensures coverage of objects at multiple scales without resizing the image multiple times.
How is Faster R-CNN different from YOLO or SSD?
Faster R-CNN is a two-stage detector: the RPN first generates object proposals, then a second stage classifies and refines each proposal. This two-stage process improves accuracy, especially on small objects, at the cost of speed. YOLO and SSD are single-stage detectors that predict boxes and classes directly in one forward pass, achieving higher throughput but typically lower accuracy on challenging benchmarks, particularly for small or densely packed objects.
Do I need a GPU to use Faster R-CNN?
In practice, yes. The shared convolutional backbone (e.g. ResNet-50) involves hundreds of millions of multiply-accumulate operations per image. On a modern GPU, inference takes 80–200 ms per image; on CPU this rises to several seconds per image, which is impractical for most applications. Training requires a GPU with at least 8 GB of VRAM for typical batch sizes.
How many annotated images are needed to train Faster R-CNN?
With a backbone pretrained on ImageNet, fine-tuning can work with a few thousand annotated images per class in many practical domains. The PASCAL VOC benchmark uses roughly 5,000 training images across 20 classes. For highly specialised domains (e.g. rare medical findings), transfer learning from a related pretrained detector and data augmentation can extend usability to a few hundred images, though performance will be lower and should be validated rigorously.
Sources
- Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. Advances in Neural Information Processing Systems (NeurIPS), 28, 91–99. link ↗
- Ren, S., He, K., Girshick, R., & Sun, J. (2017). Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6), 1137–1149. DOI: 10.1109/TPAMI.2016.2577031 ↗
- Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning (Ch. 9: Convolutional Networks). MIT Press. ISBN: 978-0-262-03561-3
How to cite this page
ScholarGate. (2026, June 3). Faster Region-based Convolutional Neural Network. ScholarGate. https://scholargate.app/en/deep-learning/faster-r-cnn
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
Compare side by side →