I'm trying to convert calibration matrix to Blender's camera as the reference document in here: http://ksimek.github.io/2012/08/14/decompose/

It looks Blender camera uses bot-left corner as origin in image space.

Is there a good way to convert this?

Here is my solution.


    scene = bpy.context.scene
render = bpy.context.scene.render

# view port 

render.engine = 'CYCLES'
render.filepath = '//'
render.resolution_x = imgWidth
render.resolution_y = imgHeight
render.resolution_percentage = 100

scene.cycles.device = 'GPU'

# K 
camdata = bpy.data.cameras.new('cameraData')
camdata.lens_unit = 'FOV'

f = float(K[0,0])
fx = float(K[0,0])
fy = float(K[1,1])
ppx = float(K[0,2])
ppy = float(K[1,2])

maxdim = max(render.resolution_x,render.resolution_y) 

#camdata.angle = 2*math.atan(0.5*maxdim/f)

# the unit of shiftXY is FOV unit (Lens Shift)
camdata.shift_x = (ppx - render.resolution_x/2.0)/maxdim
camdata.shift_y = (ppy- render.resolution_y/2.0)/maxdim

camdata.dof_distance = 0.0
camdata.clip_end = 1000.0

cam = bpy.data.objects.new('camera', camdata)
# flip axis to top-left

Extrinsic R|t

Assume all your frame shares using same camera and I add object axis to handle camera rotation. It seems difficult to directly insert R|t matrix into camera object. Also, please note that the initial order of the rotation and location parameter does matter.

    # extrinsic matrix
    all_wrtx,all_fid = top.loadALLWxyzRt(camDIR);


D = bpy.data

for o in D.objects:


for i in range(0,len(all_fid)):

    fid = all_fid[i];



    # set R

    # set Rt

  • $\begingroup$ At the end of the second paragraph there is a link to an article which describes how to convert the camera coordinates to be used with OpenGL. I haven't read it completely, but Blender's camera coordinates are pretty much the same. Have you had a look at that article? $\endgroup$
    – maddin45
    Jun 14 '14 at 9:29
  • $\begingroup$ Embed camera matrix into OpenGL is easy, because it allows you to directly touch projection matrix. However, this not allowed in the Blender. Also, if you touch the camera data, such as bpy.object.data.matrix_world, Blender would do some underly updating which make this coordinate turning become complicated. $\endgroup$
    – tirth
    Jun 14 '14 at 12:07
  • $\begingroup$ @tirth Could you also add your code (including parameters) I'm really curious. $\endgroup$
    – stacker
    Jun 14 '14 at 18:31
  • $\begingroup$ @stacker the difficulty of this problem as I aware is doing coordinate turning on camera would interfere the extrinsic setting of camera object in Blender. The one solution I found is adding another axis object to simulate camera moving. $\endgroup$
    – tirth
    Jun 16 '14 at 0:35
  • 1
    $\begingroup$ @tirth Thanks for sharing your solution, it's perfectly ok to post it as an answer and accept it. $\endgroup$
    – stacker
    Jun 16 '14 at 5:32

I've only one time used singular value decomposition with OpenCV

As I understand your question manipulating x0, y0 of the camera's intrinsic parameters (as interactivly possible in the 2nd page of your link) is similar to

import bpy

cam = bpy.data.cameras['Camera']
cam.shift_x += .5
cam.shift_y += .5

If this doesn't help you could try to offset the cameras position:

cam = bpy.data.objects['Camera']
cam.location.x = 0.0
cam.location.y = ...
cam.location.z = ...

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.