Fast R-CNN | ML

Last Updated : 26 Jun, 2026

Fast R-CNN is an improved object detection framework that addresses the inefficiencies of R-CNN by sharing convolutional computations across the entire image and performing region-based detection on a single feature map. This significantly reduces computation time while improving accuracy and enabling end-to-end training of the detection pipeline.

  • Processes the full image once through a CNN to generate a shared feature map for all region proposals.
  • Uses RoI pooling with softmax classification and bounding box regression for efficient and accurate detection.

Architecture

Fast R-CNN improves object detection by processing the image only once through a CNN and sharing the computed feature map across all region proposals.

max_path_sum_2_special_nodes_6
Fast R-CNN Architecture
  • Input and Region Proposals: The input image is combined with region proposals generated using Selective Search, where around 2000 proposals are created for potential object locations.
  • Feature Extraction using CNN: The entire image is passed through a CNN once to produce a shared convolutional feature map.
  • RoI Pooling Layer: Region proposals are mapped onto the feature map, and each is converted into a fixed-size feature vector using the RoI Pooling layer.
  • Classification Head: The extracted features are passed into a softmax classifier to predict the object category for each region.
  • Bounding Box Regression Head: The regression head refines the coordinates of each detected object to improve localization accuracy.

CNN Network of Fast R-CNN

Fast R-CNN uses pre-trained ImageNet models such as VGG-16, which consist of convolution and max-pooling layers for feature extraction. The architecture is modified in three main ways:

  • The network takes both the input image and region proposals as input.
  • The final pooling layer is replaced with the RoI Pooling layer to handle region proposals.
  • The final classification layers are replaced with two branches: a softmax classifier for object classification and a bounding box regressor for localization.
2056957910
VGG-16 Architecture

Region of Interest (RoI) Pooling

RoI Pooling is a key component introduced in Fast R-CNN that converts variable-sized region proposals into fixed-size feature maps. This ensures that all regions have a uniform representation before being passed to fully connected layers.

It takes two inputs:

  • A convolutional feature map obtained from the CNN (for example, 14 × 14 × 512 in VGG-16).
  • A set of region proposals represented as an N × 4 matrix, where each proposal defines the coordinates and size of a region (r, c, h, w).

RoI Pooling works by dividing each region proposal into a fixed number of equal sections and applying max pooling within each section. This produces a fixed-size output regardless of the original region size, allowing consistent feature extraction for all proposals.

region_proposal_network
Region of Interest Pooling

Training and Loss Function

Fast R-CNN uses a multi-task loss function to jointly optimize object classification and bounding box localization for each Region of Interest (RoI). During training, each RoI is associated with a ground truth class and a ground truth bounding box. The model compares these labels with the outputs of the softmax classifier and bounding box regressor to compute the final loss.

The multi-task loss is defined as:

L = L_{cls}(p, u) + \lambda \, u \, L_{loc}(t, v)

  • L_{cls} (Classification Loss): Measures how accurately the model predicts the object class using log loss.
  • L_{loc} (Localization Loss): Measures how closely the predicted bounding box matches the ground truth bounding box.
  • A balancing parameter λ controls the contribution of classification and localization losses.
  • The variable u ensures localization loss is applied only when the region contains an object (u = 1), and ignored for background regions (u = 0).

Here, Lcls is log loss and Lloc is defined as the localization loss function for bounding box regression.

Results and Performance Comparison

Fast R-CNN achieves state-of-the-art performance on the PASCAL VOC 2007, 2010, and 2012 datasets. It significantly reduces both training and inference time compared to R-CNN due to shared convolutional computation and end-to-end training.

  • Improves detection speed by eliminating repeated CNN processing for region proposals.
  • Reduces training time and computational cost while maintaining or improving accuracy (mAP).
r_cnn_vs_fast_r_cnn
Performance comparison between R-CNN and Fast R-CNN

Advantages

  • The entire image is processed once through the CNN, reducing repeated computation for each region proposal.
  • Combines classification and bounding box regression in a single unified network, improving optimization and accuracy.
  • RoI pooling and shared feature maps improve object localization compared to traditional R-CNN.

Challenges

  • It still uses Selective Search, which is slow and not learnable, limiting overall speed.
  • Although faster than R-CNN, the region proposal step prevents it from achieving real-time performance.
  • Processing thousands of RoIs per image still requires significant memory and computation.
Comment

Explore