When generating anchor boxes from a feature map, the center of each box corresponds to a spatial unit (pixel) on the map. Because the coordinate values of these centers are divided by the width and height of the feature map, they are transformed into relative positions between $$0$$ and $$1$$. This normalization guarantees that the resulting anchor box centers are uniformly distributed across the relative spatial dimensions of any input image, providing evenly spaced candidate regions for object detection.

Claude

To effectively detect objects of varying sizes, an object detection system can generate anchor boxes at multiple scales by utilizing feature maps of different spatial resolutions. Intuitively, smaller objects are more likely to appear on an image than larger ones; for example, a $$1 	imes 1$$ object can appear in more possible locations on a given image than a $$2 	imes 2$$ object. Therefore, a larger feature map provides more individual spatial units, allowing the system to uniformly sample a greater number of regions. This dense sampling is highly suitable for generating smaller anchor boxes to detect small objects. Conversely, by reducing the spatial dimensions of the feature map, the system samples fewer regions, which is ideal for placing larger anchor boxes intended to capture larger objects.

Multiscale Anchor Box Generation

Dive into Deep Learning

Uniform Spatial Distribution of Anchor Boxes

At a specific scale represented by a feature map of shape $$h 	imes w$$, a convolutional neural network generates $$hw$$ sets of anchor boxes, with each set containing $$a$$ anchor boxes sharing the same center. Once each anchor box is labeled with a ground-truth class and offset, the object detection model must predict these classes and offsets for all $$hw$$ sets of anchor boxes across the input image.

Learn Before

Related