When model is trained by scale jittering, the flying points caused by bilinearly resize dense gt, for example, sceneflow, will do harm to the boundary performance. Would it be proper to resize sparse gt or resize nearest?
We use linear interpolation since it doesn't negatively affect downstream performance, but it's worth trying nearest neighbor interpolation if you want the model to not produce flying pixels on object boundaries.
When model is trained by scale jittering, the flying points caused by bilinearly resize dense gt, for example, sceneflow, will do harm to the boundary performance. Would it be proper to resize sparse gt or resize nearest?