How can I get the intrinsic matrix of the agent camera

[xiaobaishu0097]

Hi,

This is probably a very amateur question but is it possible to get the intrinsic matrix of the agent camera OR can I obtain point cloud from RGB-D image captured from ai2thor? I tried https://github.com/allenai/ai2thor/issues/110#issue-412514905, but it seems to have some compiler errors.

Thanks

[ruinianxu]

@xiaobaishu0097 Hi,

I am meeting the same problem as you. I wonder if you find the way to address it. Really appreciate your help.

Thanks

[zkytony]

Just in case anyone runs into this issue, here is my two cents. I found that it is sufficient to compute the camera intrinsic using just the field of view, which is exposed by the ai2thor API and configurable.

I was originally following #275 and #110, and tried to print out the projection matrix in unity, but it led to nowhere. If you have a method to get camera intrinsic matrix directly from unity, please share.

Below is some code for plotting point cloud given a color and depth image, using Open3D. Ai2thor version is 3.3.4.

import math
import numpy as np
import open3d as o3d
from ai2thor.controller import Controller

def to_rad(th):
    return th*math.pi / 180

def example():
    width, height = 600, 600
    fov = 90
    controller = Controller(scene="FloorPlan1",
                            width=width,
                            height=height,
                            fieldOfView=fov,
                            renderDepthImage=True)
    controller.step(action="RotateLeft", degrees=45)
    event = controller.step(action="Pass")

    # Convert fov to focal length
    focal_length = 0.5 * width * math.tan(to_rad(fov/2))

    # camera intrinsics
    fx, fy, cx, cy = (focal_length, focal_length, width/2, height/2)

    # Obtain point cloud
    color = o3d.geometry.Image(event.frame.astype(np.uint8))
    d = event.depth_frame
    d /= np.max(d)
    depth = o3d.geometry.Image(d)
    rgbd = o3d.geometry.RGBDImage.create_from_color_and_depth(color, depth,
                                                              depth_scale=1.0,
                                                              depth_trunc=0.7,
                                                              convert_rgb_to_intensity=False)
    intrinsic = o3d.camera.PinholeCameraIntrinsic(width, height, fx, fy, cx, cy)
    pcd = o3d.geometry.PointCloud.create_from_rgbd_image(rgbd, intrinsic)
    pcd.transform([[1, 0, 0, 0],
                   [0, -1, 0, 0],
                   [0, 0, -1, 0],
                   [0, 0, 0, 1]])
    o3d.visualization.draw_geometries([pcd])

if __name__ == "__main__":
    example()

Output looks like:

I haven't tested, but it is promising, based on the quality of this point cloud, you can obtain 3D coordinates in the world frame from 2D pixels.

[ruinianxu]

I successfully modified the source code and let the agent print projection matrix of the camera. I've documented everything in my forked version. Under the test folder, there is a readme file documents everything from installing unity in ubuntu to compling and building the source code. Hope it can help someone who need to get the projection matrix of camera.

[GuardianWang]

focal_length = 0.5 * width * math.tan(to_rad(fov/2))

Should it be

focal_length = 0.5 * width / math.tan(to_rad(fov/2))

[LucBourrat1]

I agree with @GuardianWang. In this case it works because math.tan(to_rad(fov/2)) ~ 1 so multiplying or dividing gives the same result, but the right formula is with dividing

[ChongjianGE]

hi @ruinianxu, it seems your code should run based on a local unity hub. is that right?

[ruinianxu]

@ChongjianGE Yeah.

[ChongjianGE]

@ruinianxu Hi, ruinian,

Thank you for your assistance. I have successfully implemented the retrieval of the projection matrix in Unity Editor version 2021.03. However, I have noticed that regardless of how I move the agent/camera, the 4x4 projection matrix I obtain remains unchanged. This seems abnormal to me based on my understanding. Additionally, I would like to ask if you have any recommended methods for conveniently verifying whether the projection matrix is retrieved correctly.

[lvjihui]

i know event.metadata['agent'] can give me the position of the agent,but how can i get the relationship between the camera coordinate system and the agent coordinate system.