Open3D

그냥 Open3D rendering이 다 OpenGL convention쓰는거같은데 저 이슈의 답변은 뭘까... 

look_at 함수 참조는 이게 훌륭하다. http://learnwebgl.brown37.net/07_cameras/camera_introduction.html

 

________________________________________________________

Open3D. Camera coordinate은 convention을 따른다고 하는데 (https://github.com/isl-org/Open3D/issues/1347), 이미지렌더링은 OpenGL을 쓰기 때문에 Y,Z가 flip되어있음. 그래서 y,z 부호바꿔줘야함. 정확히 말하면, 나는 현재 Open3D로 geometry 그리고 poll_events()로 screen에 렌더링하고 그걸 capture_screen_float_buffer(False)로 capture하고 있는데, 그러다보니 우분투렌더링은 OpenGL을 써서 camera coordinate이 다름. 스크린렌더링안하고 버퍼에서 바로 캡쳐하면 문제없을듯.

if __name__ == "__main__":
    sample_data = o3d.data.DemoCustomVisualization()
    pcd_flipped = o3d.io.read_point_cloud(sample_data.point_cloud_path)
    # Flip it, otherwise the pointcloud will be upside down
    pcd_flipped.transform([[1, 0, 0, 0], [0, -1, 0, 0], [0, 0, -1, 0],
                           [0, 0, 0, 1]])

    print("1. Customized visualization to mimic DrawGeometry")
    custom_draw_geometry(pcd_flipped)

    print("2. Changing field of view")
    custom_draw_geometry_with_custom_fov(pcd_flipped, 90.0)
    custom_draw_geometry_with_custom_fov(pcd_flipped, -90.0)

    print("3. Customized visualization with a rotating view")
    custom_draw_geometry_with_rotation(pcd_flipped)

    print("4. Customized visualization showing normal rendering")
    custom_draw_geometry_load_option(pcd_flipped,

Solutions
To get the correct camera coordinates of objects and to project objects to the image plane correctly, one should invert the y and z values of objects' coordinates in the OpenGL camera coordinate system. Physically, this means rotating the objects 180 degrees around the x-axis.
Alternatively, as Shubham suggested, one can do below

# world_T_cam: 4x4 transformation matrix that defines a camera pose in the world coordinate system
# In other words, the inverse of `world_T_cam` transforms a 3D point in the world coordinate system to the camera coordinate system
world_T_cam[:, 1:3] = -world_T_cam[:, 1:3]

Or

world_T_cam[:3, :3] = world_T_cam[:3, :3] @ np.diag([1., -1., -1.])  # np.diag([1.,-1.,-1.]) is a 3D rotation matrix that defines 180 degree rotation around the x-axis; this should be multiplied on the right side of world_T_cam