Trouble using linemod

[JimmyDaSilva]

Hi, I am having trouble detecting the tutorial's coke can.

The /real_icpin_ref seems to have the coke can + a single point (See the green dot in the images). I don't really understand where this dot comes from.

[nlyubova]

Hello, Linemod requires a mesh with an origin in the center of an object (that is not a case for the coke mesh in the tutorial) that is the reason of a small shift between the meshes (in green and blue) You could try to open your mesh (in a Blender, for example) and move the origin to its centermass

[JimmyDaSilva]

Thanks @nlyubova. That will probably help :)

[JimmyDaSilva]

Hello, @nlyubova , I shitfed the mesh so that the origin is set in the center of mass. Also tried shifting it to the bottom of the can. In any case I get the same result. It contains two issues :

• green cloud /real_icpin_model is shifted (I don't really care, it's just for visualization)
• coke can recognized orientation is wrong

I have a few other problems:

• I had to remove the robot from the depth image because the coke can was half of the time found on the robot's cloud (as showed on my first post)
• I tried to use the KinectV2 but it doesn't work, neither for sd or hd images. I saw you talked about it on the google user group. You were saying there will be problems if the image ratio is different, and that the case the Kinect2 images

Thanks for your help, Jimmy

[nlyubova]

You should shift the origin to teh center of the object (center of mass) but not to the bottom (only Tabletop required a mesh with an origin at the bottom).

The shift and wrong orientation can be due to several problems: either the origin of mesh is shifted or the image frame_id is wrong ! For the second, it can be your case. check the frame_id of you depth and RGB images and check if it corresponds to the one in the conf file (especially under the parameter depth_frame_id: '.....'

ANother thing to check is the image resolution in the training conf file. Doe sit correspond to the image resolution during recognition? AN example from the conf file: renderer_width: 320 renderer_height: 240

Regarding false-positives, try to play with thresholds; also icp thresholds

and for the KinectV2, what is the reolution of color and depth images and what is the format 8U/16U/32F?

Let me know !

[JimmyDaSilva]

@nlyubova thank you for your concern!

I shifted the mesh accordingly it indeed resolved my orientation problem. I posted a PR(https://github.com/wg-perception/ork_tutorials/pull/10) on ork_tutorials with the shifted mesh. The green real_icp_model is still shifted though, not really a problem but still.

My mouse, or my arm would still give false-positives with quite high scores (~92). How are you supposed to tweak the following params :

• th_obj_dist
• icp_dist_min
• px_match_min

How do you specify the renderer's width and height? I added your lines in training.conf but it didn't changed the renderer window's size, or the display.

Concerning the KinectV2, HD images are 1920x1080, QHD images are 960x540, SD images are 512x424

Again thanks for the help :)