search

courtois-neuromod / friends.stimuli

1 stars 1 forks source link

higher resolution videos? #6

Open yibeichan opened 7 months ago

yibeichan commented 7 months ago

@bpinsard Hello, the current stimuli have a resolution of 480x720, which is too low for some analyses I need to do. I understand that the resolution fits the fMRI screen. I was just wondering whether you have higher-resolution videos available.

bpinsard commented 7 months ago

This is the native NTSC resolution from the dvd, not the MRI screen. Even if you get a BluRay the limiting factor will be the resolution it was originally released on, this is an old TV show!! You can also see de-interlacing artifacts when there are large motion. I am curious about what analysis the resolution is limiting for? Also, in the movie10 dataset, the stimuli are from BluRay so higher res.

yibeichan commented 7 months ago

i see i see, i want to identify each character's face (not face detection, which is fine). I cut the movie into pictures by frame (and second), which caused a lot of blurring that made this identification task harder. Would you happen to have some better suggestions? Does your lab try face identification on this?

bpinsard commented 7 months ago

I understand, this is indeed a very interesting type of annotations, we have not attempted that yet. Are you using something like https://github.com/psychoinformaticslab/pliers or other existing models? Tagging @MarieStLaurent here, as she has been working on other type of annotations of video stimuli.

yibeichan commented 7 months ago

I didn't use pliers but tried fine-tuning yolov9 with annotated pictures. Fine-tuning performance was okay, but the results were not satisfying... then I came up with some engineering solutions as much as I could think about. Still not as good as expected. I talked to Alejandro about it yesterday; he also said that it would be hard to get higher accuracy (given what I want).