About the downsample ratio

#3
by lucasjin - opened

From the code, pixel shuffle actually didn't changed any hw dimension, how does it ablt to downsample?

OpenGVLab org

Pixel Shuffle rearranges an image’s height and width into the channel dimension for downsampling.

Key Steps:

  1. Input Feature Map: Starts with dimensions ((H, W, C)).
  2. Rearranging Dimensions: Converts spatial dimensions into channels. For example, ((2H, 2W, C)) becomes ((H, W, 4C)).

Outcome:

  • Reduced Spatial Dimensions: Height and width decrease.
  • Increased Channel Depth: Number of channels increases.

This effectively downscales the image by packing spatial details into more channels.

czczup changed discussion status to closed

Hi, does it better compare with Resampler? Or convolution?

Sign up or log in to comment