— Object Detection in 20 Years: A Survey, 23

1. 2D Convolution

2. 3D Convolution

parameters: |channels| * 553, i.e. 55|input channels|*|output channels|

2.1 3D 1 × 1 convolutions

also called pointwise convolutions

3. Spatial separable convolution / Flattened convolution

The number of parameters in flattened convolution decreases from XYC to X + Y + C, where C is the number of input planes, X and Y denote filter width and height.

Additional factorization in spatial dimension such as in [16, 31] does not save much additional computation as very little computation is spent in depthwise convolutions.

— from mobilenet v1

如果是large kernel呢？二者结合呢？