Question about Depth Estimation

[AlexanderJGomez]

Hello, I am currently using depth estimation with object detection to calculate the lat, longs of objects in a space and was hoping to get a sanity check on the logic of this process.

With the following device settings:

        self.nodes.stereo.setRuntimeModeSwitch(True)
        self.nodes.stereo.setLeftRightCheck(True)
        self.nodes.stereo.setSubpixel(True)

        self.nodes.stereo.setDepthAlign(dai.CameraBoardSocket.RGB)

I take the bounding box from the color camera and scale it to the depth frame. Then, I use the scaled bounding box to calculate the average depth over that region in the depth frame. Using that average depth, I approximate x, y, z coordinates, similar to how it is done here.

Then, I calculate the absolute distance from the center of our camera to the point and the angle (before converting to lat long). I calculate the distance by doing the following:

    dist = math.sqrt(x**2 + z**2) # ignores height of point

Lastly, given the angle to the object and the distance away it is (along with other information about where I am and where I am facing), I calculate the object's location.

[Erol444]

Hello @AlexanderJGomez , I believe the Z is already the distance to the object, and Xand Y are then calculated from the Z.

[AlexanderJGomez]

Hi @Erol444 thanks for the response. I'm a bit confused because this example seems to indicate that z is just another dimension in a 3D cartesian space

[SzabolcsGergely]

Z is indeed just another dimension in a 3D cartesian space, the distance to the object is the Euclidean distance. sqrt(x^2+y^2+z^2)
Depth map basically returns the Z dimension, if you point the camera to a flat surface, all the values will be the same (given that surface has enough texture).