RE: SEEING MACHIN]ES LTD [AU] - 2022-07-2126 Jul 2022 00:42
Apologies for the delay, I was away so didn't see the bat signal on the Canberra horizon. I haven't had time to digest the whole patent - it isn't much fun reading it on a mobile, so just a few points for now.
When I was growing up, we thought by the new millenium, we would all be wearing shiny suits and be driven around in flying cars by our robots. But reality got in the way and now we need to babysit our robots which are only really capable of steering and and controlling the speed - assume they can work out where the road lines are and they correctly guess the behaviour of EVERYTHING on the road. Of course babysitting is boring, you have to stay awake, and you can't even have a drink or watch a film in the car - unless you cheat., since we are still the responsible adult in the room and we can either die or face legal repercussions for not supervising the driving properly, or doing it ourselves the old fashioned way we need help behaving
So how to Seeing Machines make sure the humans have skin in the game - well they look for our skin! You see, it is possible to fool a camera by showing it a picture, so how do we avoid needing 2 cameras (and 3-d processing) well we can modify the lighting. I will pick one "simple" option. Imagine a flat light source that magically contains vertical stripes like pinstripes. If you project this light onto a sheet of paper, you will still get straight lines. View the paper from an angle and it still has straight lines. Now imagine a face, the lines will work like contours as long as they are viewed at an angle from the illumination. Change the light source so that the lines are horizontal (assuming the camera is to the side) and the effect almost vanishes in line with the camera, but will still be visible as you move up or down.
OK, next step, use a grid of light spots. On a flat sheet of paper, the image of the spots might be 100 pixels apart. Now put a football in the light instead of a head. Viewed slightly to one side, we see the central square is probably still around 100 pixels, but as you move away, either the width or thickness of the squares becomes larger and the shapes are no longer square.
Now you can draw a graph these distances. On the flat sheet of paper, everything is around 100 pixels, so there is a huge spike. Change to a ball and the range of measurements varies greatly - but smoothly, and moving the ball or rotating it doesn't change the pattern much. OK now use a face and the complex "topology" gives a far wider spread of values. Flat foreheads, chins and brows may be mostly flat, but as the head turns, these will distort the squares and result in longer or shorter distances.
That is called a Fast Fourier Transform and is relatively easy to develop and process, but it can tell a flat picture from a curved picture, a ball or a head quite easily, if you can feed it a "spatial pattern".
I didn't get to "skin" or speckles yet, but I am out of characters, so that is a story fo