search

NVlabs / HANDAL

HANDAL Dataset and Pipeline
https://nvlabs.github.io/HANDAL/
Other
57 stars 1 forks source link

performance #12

Open EladUnlimited opened 2 weeks ago

EladUnlimited commented 2 weeks ago

hi all I like to ask two questions please: edit: I have found a mistake with my calculation so question 1 is not relevant

  1. i did use the demo in the git and it works pretty good (as shown in fig 1). in the other hand, when i use costume data the results are not very accurate (as shown in fig 1 & 2). do you have any idiea how to improve it or a general guidence to how to perform the prosses for
    a better results? note: the resolution for the costume data is 428x850 (the original
    drill video is 1920x1080) 2) i did use Cutie as recommended. for the 428x850 resolution it works very good but for the original resolution (1920x1080) Cutie is not able to propagate the segmentation over to the following frames. i have tried the original drill video as well but Cutie didn't make it to propagate. can you please share how did you labeled the original video?

fig 1 - test

fig 2- test2

fig 3 - test3

andrewyguo commented 1 week ago

Hello @EladUnlimited,

Could you elaborate on your issue with Cutie? Are you finding that the masks are not propagating when you run at higher resolutions?

For our purposes, we found that masks that were 960x540 were enough to cleanly segment the object. This will run faster since it is processing only 1/4 of the pixels compared to 1920x1080.

Do you need masks at 1920x1080 for a specific purpose?

EladUnlimited commented 6 days ago

Hello @EladUnlimited,

Could you elaborate on your issue with Cutie? Are you finding that the masks are not propagating when you run at higher resolutions?

For our purposes, we found that masks that were 960x540 were enough to cleanly segment the object. This will run faster since it is processing only 1/4 of the pixels compared to 1920x1080.

Do you need masks at 1920x1080 for a specific purpose?

Hi Regarding your firs question, yes. When I am using the 1920x1080 videos the masks are not propagating.

I don't specifically need to have 1920x1080, just tried to replicate your work to see if I having some problems with the process or is it just the videos that are too large. I find that any way I can't propagate the masks on the original videos.

So in the example you are providing is videos being use are the original one after resize for 960x540? my question (with more details) is, did you just resize the videos for the purpose of the segmentation with Cutie (and than running on your project with the original 1920x1080 resolution and the masks resize again to 1920x1080) or is the entire process (Cutie and HANDAL) is using the 960x540 resolution?

Thank you!