ROS Resources: Documentation | Support | Discussion Forum | Index | Service Status | Q&A
Ask Your Question

Having problems with TOD object detection

asked 2013-08-07 22:51:37 -0500

bit-pirate gravatar image

updated 2013-08-07 23:34:53 -0500

Hi there! I have been cooking in the object recognition kitchen for a bit and so far I succeeded at creating my own models, uploading and training them using TOD.

Unfortunately I can't get the detection to work - i.e. zero poses found. The output looks like this:

***Starting object: 0
5 keypoints in 6 matches
done filling
RANSAC done with 0 inliers
***Starting object: 1
12 keypoints in 51 matches
done filling
RANSAC done with 0 inliers
***Starting object: 2
2 keypoints in 4 matches
done filling
RANSAC done with 0 inliers
***Starting object: 3
4 keypoints in 4 matches
done filling
RANSAC done with 0 inliers
***Starting object: 4
2 keypoints in 2 matches
done filling
RANSAC done with 0 inliers
***Starting object: 6
6 keypoints in 24 matches
done filling
RANSAC done with 0 inliers
********************* found 0 poses

Comparing this with the output shown in the tutorial I noticed that I get very few keypoints and matches and RANSAC returns 0 inliers.

Any guess what is going wrong here?

Side note: I'm using Groovy with a source install on Ubuntu Precise 64bit and my sensor is the ASUS Xtion Pro Live.

Update: I can report a tiny success! I switched to the Kinect, activated the high res mode (SXGA for both depth and rgb), captured and trained the new object and I'm getting a found 1 poses now.

***Starting object: 0
5 keypoints in 5 matches
done filling
RANSAC done with 0 inliers
***Starting object: 1
85 keypoints in 1103 matches
done filling
RANSAC done with 52 inliers
RANSAC done with 5 inliers

Unfortunately, I don't get any (valid) output on the rostopics - markers, poses, recognized_object array are all empty. Is this a bug?

So, my question now turns into this one: Does TOD only work with high res models and input?

edit retag flag offensive close merge delete


I also was not able get any poses using vga @ 30 fps. bit-pirate: When I used --res sxga and --fps 15 on the object_recognition_capture pkg 'capture' executable, the program locks up. I am using: -a Kinect -set the dynamic reconfiguration(through rqt) to sxga @15fps -using groovy/64bitUbuntu 12.04

bk gravatar image bk  ( 2013-10-04 09:32:13 -0500 )edit

Hm, hard to say. I haven't been using TOD for a while now, since the reimplemented tabletop object detector is sufficient for my needs. However I did the tests described above with the exact same specs as you. Could you provide more details about where/how the program get's stuck?

bit-pirate gravatar image bit-pirate  ( 2013-10-06 23:18:58 -0500 )edit

The program hangs at: "Setting sync off". One possible difference is that my depth resolution seems to stay at vga@30fps. Using rqt and dynamic reconfigure, I change the depth resolution to sxga @ 15fps but when I exit out of rqt and enter it again the depth goes back to vga@30fps.

bk gravatar image bk  ( 2013-10-07 10:33:12 -0500 )edit

I am also getting a similar output. I am using kinect sensor, groovy, 64 bit. Did you find any solution?

gautam2410 gravatar image gautam2410  ( 2013-11-26 18:05:09 -0500 )edit

1 Answer

Sort by ยป oldest newest most voted

answered 2014-04-15 00:58:23 -0500

Related update, ORK using TOD and tabletop kind of works, I wrote a post with all the steps I followed to get some detections with them: Using ubuntu 12.04, ROS Hydro, last ORK from sources, 32b and 64b tested.

edit flag offensive delete link more

Your Answer

Please start posting anonymously - your entry will be published after you log in or create a new account.

Add Answer

Question Tools



Asked: 2013-08-07 22:51:37 -0500

Seen: 1,070 times

Last updated: Aug 07 '13