First of all thanks for sharing this great work! I am trying to train on my own videos so I need to extract motion vectors (MVs) first, which I'm trying to do by calling your CoViAR's loadmethod (a big shout out for that work as well!). I have two questions:
1) The load method, by default, computes the MV w.r.t. the previous frame, correct? So to get the "after" flows as specified in the sample training/eval data, I can call load for a second time on the video with its frames reversed, correct?
2) By looking at read_bmv, it seems that to convert the int32 MV into uint8, you added 128 to all of its elements and possibly did another clamping to [0, 255]. I'm not very familiar with MV so could you please explain a little bit on why this will not cause loss of information?
Hi Chao-Yuan,
First of all thanks for sharing this great work! I am trying to train on my own videos so I need to extract motion vectors (MVs) first, which I'm trying to do by calling your CoViAR's
load
method (a big shout out for that work as well!). I have two questions:1) The
load
method, by default, computes the MV w.r.t. the previous frame, correct? So to get the "after" flows as specified in the sample training/eval data, I can callload
for a second time on the video with its frames reversed, correct?2) By looking at
read_bmv
, it seems that to convert theint32
MV intouint8
, you added 128 to all of its elements and possibly did another clamping to [0, 255]. I'm not very familiar with MV so could you please explain a little bit on why this will not cause loss of information?Many thanks in advance.