search

tum-vision / lsd_slam

LSD-SLAM
GNU General Public License v3.0
2.58k stars 1.23k forks source link

TRACKING LOST error while using dataset SLAM #259

Open ulkesh11 opened 7 years ago

ulkesh11 commented 7 years ago

I am using Iphone 6s camera to take a video and then extract each frame fro the video to form the dataset. My calibration file is as follows.

0.8609 1.1474 0.4999 0.4966 0.062 1920 1080 crop 640 480

After running the following coomand: rosrun lsd_slam_core dataset _files:=/home/vk/Downloads/my_data/images2 _hz:=30 _calib:=/home/vk/catkin_ws/src/lsd_slam/lsd_slam_core/calib/iphone_6s_calib.cfg

I get the following error: Reading Calibration from file /home/vk/catkin_ws/src/lsd_slam/lsd_slam_core/calib/iphone_6s_calib.cfg ... found! found ATAN camera model, building rectifier. Input resolution: 1920 1080 In: 0.860900 1.147400 0.499900 0.496600 0.062000 Out: Crop Output resolution: 640 480 new K: 550.914551 550.794739 319.769287 238.145020 old K: 1652.927979 1239.192017 959.308044 535.828003 Prepped Warp matrices Started mapping thread! Started constraint search thread! Started optimization thread found 2131 image files in folder /home/vk/Downloads/my_table/images2! Doing Random initialization! started image display thread! Done Random initialization! warning: reciprocal tracking on new frame failed badly, added odometry edge (Hacky). warning: reciprocal tracking on new frame failed badly, added odometry edge (Hacky). TRACKING LOST for frame 302 (5.27% good Points, which is 29.07% of available points, NOT DIVERGED)!

Can someone please tell me where am I going wrong?

Thanks!

FirefoxMetzger commented 7 years ago

It's hard to tell from this. TRACKING LOST indecates, that there aren't enough keypoints visible in the active frame (only ~5%). Meaning that LSD-SLAM can't tell where the camera is with respect to the active keyframe. Usually this should start the relocalizer, but it keeps failing for me, too.

You could try minimizing running into this issue by being even more mindfull of rotations and translations, thats at least what worked for me.

ulkesh11 commented 7 years ago

@FirefoxMetzger
Thanks a lot for replying. Here is my test video. Can you please suggest me any improvements on it?

sinha-kirti commented 7 years ago

Hi, I believe the problem is due to the translation motion in initial part of the video. I was facing the same problem because lsd-slam uses rotation for initial approximations. Try with video having only rotational motion in the beginning of the video for about 30 -40 frames. Also, you can try with prerectified images instead of directly feeding the images captured in video.

On Apr 19, 2017 4:26 AM, "vacky11" notifications@github.com wrote:

@FirefoxMetzger Thanks a lot for replying. Here is my test video. Can you please suggest me any improvements on it?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

FirefoxMetzger commented 7 years ago

I don't see anything obviously wrong with it. Assuming 30Hz you diverge around 10 sec, which mainly shows the laptop.

While the laptop has a pretty slick design, LSD-SLAM works on image gradients (corners and edges). Looking at it from that perspective, the laptop might be a poor choice as there is no potential for keypoints on the display and the keyboard area is pretty smooth as well (esp. at 640 x 480). You could try changing the object or show some image on the screen to give LSD-SLAM more change to find keypoints on it. Edit: same goes for the bottle of lotion. It's essentially a white block (with some blurry part in it) at 640x480.

Also you are doing offline SLAM. Try setting the framerate option to 0 Hz and increase the output resolution. This will not be real time anymore, but should give you more information about potential problems,