SpaceToDepth

SpaceToDepth is a data rearrangement operation used in computer vision and deep learning. It converts spatial information into the channel dimension by moving blocks of pixels from the height and width into the depth (channel) dimension, thereby reducing spatial resolution while increasing the number of channels.

Formally, for an input tensor with shape (N, H, W, C) and a block size r, where H and W are divisible by r, the output has shape (N, H/r, W/r, C*r*r). Each r-by-r neighborhood is unrolled into the depth dimension, so the information in a local spatial area is preserved but stored across more channels.

SpaceToDepth is the inverse operation of DepthToSpace. It is used to reformat feature maps to enable efficient processing, such as applying convolutions after downsampling or fusing multi-scale information. It appears in various network architectures and frameworks; for example, TensorFlow provides an implementation as tf.space_to_depth, and the operation is conceptually similar to the reorganization (reorg) layer used in some YOLO models.

Advantages of SpaceToDepth include the ability to downsample spatially while preserving information in the depth dimension, which can facilitate certain network designs and memory access patterns. Limitations include the requirement that height and width be divisible by the block size and the increase in channel dimensionality, which can affect subsequent layer configurations and memory usage.

See also: DepthToSpace, Reorg layer.