In a physics pixel frame, there could be invisible or overlapped interfaces and pixels in depth. How to make an ai model know an invisible pixel has visual info (RGB or color) but the frame cant see it in training data and also to make the ai model know what value of the invisible pixel to estimate or predict in reference exactly? Obviously, no value or RGB(0,0,0) both are not good choices.
So I suggest -1 as the visual info for invisible pixel in physics pixel frame, like RGB(-1,-1,-1), which means unknown or invisible visual values.
And there could be a full perspective view of all crucial objects in a physics pixel frame of a training visual flow, which includes all interfaces of all the crucial objects, which can make the ai model learn to understand/approximate the spatial shape/space from the visual better in training and then to estimate and predict the spatial shape/space from visual more accurately and efficiently.
Be First to Comment