https://thefoxofsky.github.io/project_pages/ddf

Three missions

1. Classification: ResNet50 / 101 get improved by 1.9% and 1.3% on the top-1 accuracy, while their computational costs are reduced by nearly half.
2. detection
3. joint upsampling

Dynamic Filter

Left: dynamic convolution, an “example” STN; Right: dynamic filtering. It is also a kind of “residual” network exploring multiplication instead of addition. From Dynamic Filter Networks, nips16.

Spatial Transform Networks (STN) is the first.

2 properties of convolution

Content-agnostic

1. a spatially shared filter may not be optimal to capture features across different image regions [52,42].
2. Once a CNN is trained, the same convolution filters are used across different images (for instance images taken in daylight and at night).
3. In short, standard convolution filters are shared across images and pixels, leading to sub-optimal feature learning.

Computation-heavy or memory-heavy

Other dynamic filters

Experiments