Feature pyramid network

As explained in Deep Learning with Python, Third Edition

An architecture that merges multi-scale feature maps from different depths to support detection of objects at multiple sizes.

A feature pyramid network (FPN) helps an object detector recognize objects with different pixel footprints. Final ConvNet layers provide semantically meaningful object representations but low-resolution maps; early layers preserve high-resolution details but mostly represent simple features such as edges and curves. section 12.3

RetinaNet inference

A RetinaNet preset exposes the pattern directly: detector = keras_hub.models.ObjectDetector.from_preset("retinanet_resnet50_fpn_v2_coco", bounding_box_format="rel_xywh"). Calling detector.predict(image) returns boxes, confidence scores, labels, and the number of detections. The detector can then draw each predicted box and its COCO label on the image. section 12.3

The construction follows a practical rule:

• Upsample final ConvNet features progressively with Conv2DTranspose layers. section 12.3
• Add lateral connections by summing each upsampled map with the same-sized map from the original ConvNet. section 12.3
• Use the resulting semantically meaningful maps at multiple scales to improve detection of small and large objects. section 12.3 section 12.4

Figure 12.8 A feature pyramid network creates semantically interesting feature maps at different scales.

Use the architecture sketch to trace how low-resolution semantic features and high-resolution localized features form a multi-scale representation. section 12.3

How Deep Learning with Python, Third Edition covers it

How deep it goes

Explains upsampling and lateral summation, but stops short of mathematical derivation or detailed layer design.

Examples

Uses RetinaNet COCO inference on a Pointillist painting, with predicted boxes, labels, confidence scores, and plotted detections.

What it assumes

Assumes familiarity with ConvNet layers and feature maps while clarifying why early and late representations serve different detection needs.

What you can do after

Apply it via KerasHub by loading a RetinaNet preset, predicting detections, and visualizing labeled bounding boxes.

How it connects

Connects object detection, ConvNets, RetinaNet, YOLO, single-stage detectors, and Conv2DTranspose layers.

liveBookmanning.com

Read Deep Learning with Python, Third Edition on liveBook Get the book at manning.com

Also explained in

• 

  liveBookmanning.com
  Deep Learning with R, Third EditionA feature pyramid network is a neural-network architecture that combines semantically rich low-resolution features with spatially detailed high-resolution features at multiple scales.

Compare how 2 books explain feature pyramid network →

Related concepts

• RetinaNetused by · RetinaNet is a one-stage object-detection architecture that uses convolutional feature pyramids with separate classification and bounding-box regression heads.
• feature mapuses · A feature map is a spatially organized array of neural-network activations representing the presence or strength of learned features at different locations.
• Conv2DTransposeuses · Conv2DTranspose is a neural-network layer that applies a learned two-dimensional transposed convolution to increase the spatial resolution of feature maps.
• upsamplingimplements · Upsampling is a process that increases the number, density, or spatial or temporal resolution of samples or values in a dataset or representation by adding, duplicating, or estimating them.

This explainer was generated by Manning's knowledge graph from the text of Deep Learning with Python, Third Edition. The section links point to the parts of the book it draws on.