search

ltkong218 / FastFlowNet

FastFlowNet: A Lightweight Network for Fast Optical Flow Estimation (ICRA 2021)
MIT License
259 stars 42 forks source link

Question about warp layer #4

Open wmn931201 opened 3 years ago

wmn931201 commented 3 years ago

Hi, Thanks for your great work. I am confused why flow6_up is multiplied by 0.625 here. eg: f25_w = self.warp(f25, flow6_up*0.625)

ltkong218 commented 3 years ago

Thank you for your appreciation. Let height and width of original input image to be H and W, flow6_up has spatial shape of (H / 32, W / 32), while the flow field of flow6 and flow6_up is supervised with ground truth flow divided by 20. So before we use it to warp target feature in pyramid level 5, we should multiply flow6_up by a scale factor of 20 / 32 = 0.625 to make it represent true displacement.