How to get camera relative pose

[hcv1027]

I'm trying to use camera's ground truth pose to replace PoseNet-sfm's posenet during training.
I use below way to compute camera's odometry for verification, and it looks very strange.
Here is the way I use to compute camera's odometry for a scene (A series of temporal continuous samples).

    scene_idx = 14
    size = 50
    x = []
    y = []
    z = []
    T_cam_odom_gt = None
    for i in range(size - 1):
        datum_0 = dataset.get_datum_data(scene_idx, i, 'camera_01')
        datum_1 = dataset.get_datum_data(scene_idx, i + 1, 'camera_01')
        T_c0_w = datum_0['pose'].inverse()
        T_w_c1 = datum_1['pose']
        T_c0_c1 = T_c0_w.__mul__(T_w_c1)
        if i == 0:
            T_cam_odom_gt = T_c0_c1.copy()
        else:
            T_cam_odom_gt = T_cam_odom_gt.__mul__(T_c0_c1)
        x.append(T_cam_odom_gt.translation[0])
        y.append(T_cam_odom_gt.translation[1])
        z.append(T_cam_odom_gt.translation[2])
            
    plt.scatter(x, z, marker='.')
    plt.savefig('ddad_pose.png')

I test it with scene 000014, since the car turned left at the crossroad, I expect the camera's odometry on x-z plane should something like this one:

But the output is this one:

Can you tell me what is the meaning of the pose get from datum_data['pose']?
Is it a camera's pose with respect to the world frame?

[RaresAmbrus]

Hi, sorry about these issues. Have you had any luck? For one of our papers we used the extrinsics field to compute the relative transformation between sensors, but it's possible that the odometry i.e. the pose datum is not accurate for some sequences. I'll try to see what I can find out about that.

[GANWANSHUI]

Hi @RaresAmbrus , I am also wondering whether the pose by @hcv1027 calculation could be served as the GT pose. Does it reliable? Thanks a lot!