Read .pfm to numpy and conversion from pixel to world

[zaksel]

Hi,
i am using a rc_visard 65m and this library to stream disparity images via GenICam.
I read the Image from the .pfm-File (e.g. image_946691111.740987800_Disparity_1_0.zip) with this snippet (abducted from here: https://lmb.informatik.uni-freiburg.de/resources/datasets/IO.py):

            file = open(path_to_pfm, 'rb')
            #skip metadata
            file.readline().rstrip()
            file.readline().rstrip()
            file.readline().rstrip()
            #read into numpy array Big Endian float
            data = np.fromfile(file, dtype='>f')
            #reshape and flip
            data = np.reshape(data, (480, 640))
            data = np.flipud(data)

And get this Image:

I got the cam parameters via gc_info serial_no?Scan3dControl and they are:

self.params = {'FocalLengthFactor': 0.843637, 'Scan3dOutputMode': 'DisparityC', 'Scan3dDistanceUnit': 'Pixel', 'Scan3dCoordinateOffset': 0.0, 'Scan3dInvalidDataValue': 0.0, 'Scan3dCoordinateScale': 0.0625, 'Scan3dInvalidDataFlag': 1.0, 'Scan3dFocalLength': 539.927, 'Scan3dBaseline': 0.0649325, 'Scan3dPrincipalPointU': 320.0, 'Scan3dPrincipalPointV': 240.0, 'Baseline': 0.0649325}

But when I want to calculate the pixel [240, 320] in real world coordinates like here:
https://doc.rc-visard.com/latest/en/gigevision.html#image-stream-conversions

        p_x = (x-self.params['Scan3dPrincipalPointU']) * self.params['Scan3dBaseline']/(dik*self.params['Scan3dCoordinateScale'])
        p_y = (y-self.params['Scan3dPrincipalPointV']) * self.params['Scan3dBaseline']/(dik*self.params['Scan3dCoordinateScale'])
        p_z = self.params['Scan3dFocalLength'] * self.params['Scan3dBaseline']/(dik*self.params['Scan3dCoordinateScale'])

The distance for an Object that is 80cm away is p_z=7.6m.

I think i've made a mistake reading the disparity Image from file, because my Image contains very small values. More like uint8 than uint16. And so dik is small and p_z is following by that too big.

I have tried to change dtype='>u2' when reading the .pfm but then i cant shape the Image to the correct size because i get twice as much values (614400 instead of 307200)?

Anybody knows where I have made a mistake or succeeded with getting the conversion from pixel to world done?

Thanks for your help
derEskimo

[flixr]

The Scan3dCoordinateScale is only needed when converting the 16bit image (as in the GenICam image buffer) to a floating point disparity value in pixels.

Since you are using the cli tool gc_stream which already saves the disparity image in float format (*.pfm), you alreay have the floating point value in pixels and the scaling should not be applied again.

[zaksel]

Ahh alright. Thank you! So i do the conversion like that:

p_x = (x-self.params['Scan3dPrincipalPointU']) * self.params['Scan3dBaseline']/dik
p_y = (y-self.params['Scan3dPrincipalPointV']) * self.params['Scan3dBaseline']/dik
p_z = self.params['Scan3dFocalLength'] * self.params['Scan3dBaseline']/dik