search

kaichen-z / DynPoint

dynpoint: dynamic neural point for view synthesis
https://arxiv.org/pdf/2310.18999
15 stars 1 forks source link

Problem in function "process_optical_flow" #4

Closed SleepEaaarly closed 7 months ago

SleepEaaarly commented 7 months ago

When running your code, I found after going through funtion "process_optical_flow" the output flow picture size didn't match the input height and width. Specifically, my input height is 288 and width is 540. The output former_1_2 is (2,1,2,288,536) and the same condition with former_2_1, which means the output flow picture is 288 536. I suspected it was changed by "model_flow" which is RAFT. I also noticed in the end of this function you were trying to resize the flow picture by function "interpolate". However, the input params of this func which are H, W, H_flow, W_flow always equal, which means it was impossible to get into this "else" branch. All in all, is there a bug in this function or how could i use 288 540 photos to go through this function?

kaichen-z commented 7 months ago

Hello, thank you for your interest. If the width or height of your input is not a multiple of 8, its dimensions will be adjusted. For example, if your input resolution is 288 x 540, you would need to interpolate it to 288 x 536 before feeding it into FlowNet. After obtaining the output, you'd resize it back to 288 x 540.

SleepEaaarly commented 7 months ago

Thanks very much for your reply! Hoping for your following code!