About Flow loss about hidden-two-stream HOT 2 CLOSED

Hi,

`layer {
name: "WarpDownsample6_10"
type: "Downsample"
bottom: "img10Norm"
bottom: "predict_flow6_0"
top: "downsampled_img10_6"
propagate_down: false
propagate_down: false
}

This part is only for downsampling. img10Norm is the normalized image at original size, downsampled_img10_6 is the downsampled image. They are the same image, just size is different. predict_flow6_0 here only provide the size information. Because we want to know how much we need to downsample the image. There is no warping involved here.

layer{
name: "FlowScale6_0"
type: "Scale"
bottom: "predict_flow6_0"
top: "FlowScale6_0"
scale_param {
filler {
type: "constant"
value: 0.625
}
bias_term : false
}
param {
lr_mult: 0
}
}`

For this part, it is for scaling the flow. Because for image at different sizes, we need to enlarge or decrease the magnitude of flow vectors before we perform warping. For example, if you have an image of 200x200 and a car moves 20 pixels, then the flow is 20. But if you downsample the image to 100x100, you need to divide your flow vector by 2 as well. Otherwise, the car will move too fast, which is wrong. There is no warping involved here.

layer{
  name: "Warp6_9"
  type: "FlowTransformer"
  bottom: "downsampled_img10_6"
  bottom: "FlowScale6_9"
  top: "Warped6_9"
}

For this part, this is the actual warping. downsampled_img10_6 is the downsampled image, FlowScale6_9 is the magnitude-reduced flow, and the output Warped6_9 is the reconstructed image. Hope this is clear.

from hidden-two-stream.

Hi,
you made it very clear, thank you so much..

from hidden-two-stream.