search

lmb-freiburg / dispnet-flownet-docker

Dockerfile and runscripts for DispNet and FlowNet1 (estimation of disparity and optical flow)
GNU General Public License v3.0
87 stars 19 forks source link

Questions about gt flow #18

Open SwagJ opened 3 years ago

SwagJ commented 3 years ago

Dear @nikolausmayer ,

Thank you for your awesome dataset. However, when I tried to used the dataset. Using the Python I/O routine, you provided on dataset website, the GT flow I loaded gives a large pixel displacement value around boundary, like 152.4762 or 152.5085(take /driving/optical_flow/15mm_focallength/scene_forwards/slow/into_future/left/OpticalFlowIntoFuture_0001_L.pfm as an example). However, visually, for image 0001.png and image 0002.png, the flow should not have that large value. 0001 0002

Is there something wrong with the my GT loading process or did I take the wrong image input?

flow load function used: readPFM from you official website. flowocc, = readPFM(path_to_file) flow value are reported is: flow_occ[:,:,2]

I am looking forward to your reply. Thank you very much.

Best,

nikolausmayer commented 3 years ago

Hi,

you are observing the magic of perspective distortion :slightly_smiling_face: These "huge" flows are truly correct. Try backwarping the second image onto the first — you should see that they match perfectly!

You are absolutely right in that "visually" these flows should not be there — but keep in mind that you do not actually see where the pixels on the image border move (they move outside the image!). If the field of view were a little larger, you could verify that they do indeed move that far.

Best. Nikolaus