Segmentation Issue

[SimonePanicucci]

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 16.04
• Python version: 3.5
• Installed using pip or ROS: pip
• Camera: Kinect v2
• GPU model (if applicable):

Describe what you are trying to do
I'm trying to get the prediction of suction and parallel jaw using the pre-trained network gathering depth image from real environment.

Describe current behavior
I'm gathering depth image from my Kinect using the method frames of Kinect2Sensor, it collects some noisy points which can be removed using inpaint method but in both cases I can't produce a good segmentation image.

Describe the expected behavior
segmented image

Describe the input images
I cut the image collected in order to have just the table under the camera and nothing else.

Describe the physical camera setup
I have the kinect camera over the table 60cm far.

Other info / logs
This is the original image coming from the camera:

and the output of depth_im.invalid_pixel_mask().inverse() is:

applaying inpaint to the original depth_im we obtain:

and the corresponding depth_im.invalid_pixel_mask().inverse() is

Can you please help me understanding the problem?
Many thanks.