search

chico2121 / bpe

MIT License
26 stars 11 forks source link

SARA dataset videos resolution #19

Open ghost opened 2 years ago

ghost commented 2 years ago

Hey ! I have some doubts related with the resolution of the images/videos used in the SARA dataset. I have the keypoints coordinates normalized, so I should multiply each one with some resolution apprx. the resolution used to train the model right ? Is this the best way to have better results ? Can you provide it ?
Thanks.

BAILOOL commented 2 years ago

I think the model is trained with the normalised coordinates. So multiplication by image width or height is only needed for visualisation purposes.

ghost commented 2 years ago

Hi. Thanks. But i guess the results differ a lot if I pass a motion projection normalized or multiplied by some (width,height). Passing the motion normalized will results in scores much higher than the other case and with a lot of false positives finds. Why so ?

ghost commented 2 years ago

Hi Alex. I'm still not completely convinced about the right way to pass the keypoints. I am passing the keypoints normalized (so, between 0 and 1) and the scores are very high. Almost all of them are in the house of 0.99 similarity. Why is this happening ? And if I multiply my keyoints for a certain number (imagine x100) the results change completely with a more distributed score values. Can you explain why this happening ? Thanks.