What is a Feature Pyramid Network?
A **Feature Pyramid Network**, or **FPN**, is an artificial neural network used for object detection in images. Specifically, it is a feature extractor that takes a single-scale image of an arbitrary size as input and outputs proportionally sized feature maps at multiple levels. This allows for the detection of objects at different scales within an image.
How Does FPN Work?
The construction of the pyramid involves a bottom-up pathway and a top-down pathway. The bottom-up pathway is the feedforward computation of a backbone ConvNet which computes a feature hierarchy consisting of feature maps at several scales with a scaling step of 2. One pyramid level is defined for each stage for feature pyramid. The output of the last layer of each stage is used as a reference set of feature maps. For ResNets, we use the feature activations output by each stage’s last residual block.
In the top-down pathway, FPN hallucinates higher resolution features by upsampling spatially coarser features from higher pyramid levels. These features are then enhanced with features from the bottom-up pathway via lateral connections. Each lateral connection merges feature maps of the same spatial size from the bottom-up pathway and the top-down pathway. The bottom-up feature map has lower-level semantics but more accurately localized activations as it was subsampled fewer times.
Why Use a Feature Pyramid Network?
A feature pyramid network is used in object detection, because objects in an image can vary in size, shape and orientation, making it difficult for traditional object detectors to accurately identify them. An FPN can identify an object regardless of these changing parameters, as it creates feature maps at different scales within an image. This multi-scale feature representation allows for faster and more accurate detection of objects in varying environments.
Feature Pyramid Networks are a valuable tool for object detection in images. They allow for the detection of objects at different scales and help to create a multi-scale feature representation which is critical for efficient and accurate object detection.
