Unclear how to run the semantic mask part along with orb_slam(e.g. mono_kitti node)

[mshong0320]

Been trying to figure out how to get the semantic part to run along with running the orb_slam2.
The only thing that's been working for me now is running a ros-node by ("rosrun ORB_SLAM2 mono_kitti Vocabulary/ORBvoc.txt Examples/Monocular/KITTI03.yaml") or running ("./Examples/Monocular/mono_kitti Vocabulary/ORBvoc.txt Examples/Monocular/KITTI0X.yaml PATH_TO_SEQUENCE").

Can someone provide some more detailed guidance to running the semantic mask part along with the orb slam2 to get the result as shown on the sample video included in this repo?

Thank you in advance!

[mthz]

hi, we implemented our method only with the stereo nodes.
please precompute the labels and put them into KITTI $kitti_dataset_dir/$sequence/dla/labelIds. (e.g. with https://github.com/ucbdrive/dla, the id/color to semantic label is defined in 'slamantic/labels/labels-cityscapes.yaml
'). unfortunatley we are currently not able to share the labeling code itself.
then use e.g. KITTI03-df.yaml for the same configuration as shown in the video.

[mshong0320]

Thank you for the reply. After precomputing the labels do I have to run both the DynaSLAM and the OrbSLAM together at the same time? I'm still confused about how to use the output from DynaSLAM as an input to the OrbSLAM2.

[humenbergerm]

Hi! DynaSLAM is a complete different method. We just used it to compare or method to the state of the art. You only need to run the stereo node examples with precomputed semantic segmentation images. Best, Martin

…

On Sat 9 Nov 2019 at 01:58, Minsung Chris Hong ***@***.***> wrote: Thank you for the reply. After precomputing the labels do I have to run both the DynaSLAM and the OrbSLAM together at the same time? I'm still confused about how to use the output from DynaSLAM as an input to the OrbSLAM2. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=ABCNPODP5WR2XBONYR56U73QSYDJTA5CNFSM4JKJWPL2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEDTZHGQ#issuecomment-552047514>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABCNPOGLWHGSC7LWDLKEM6TQSYDJTANCNFSM4JKJWPLQ> .

[mshong0320]

Hi,

Still have some issues. Right now I'm stuck on how to use the https://github.com/ucbdrive/dla to precompute the labels. Can you give me some guidance on that? I looked through the dla repo but it isn't clear where in the files precomputation of labels is happening. Also I downloaded all 6 vkitti files and am trying to run vkitti_create_extrinsic.py choosing the --data directory as ./vkitti_1.3.1_extrinsicsgt/. I chose that directory as none of the other ones contains any extrinsic files. None of the vkitti dataset folders that I downloaded from Naverlabs Europe has extrinsic.txt files in them. Am I missing something here? Also on line 92 of vkitti_create_extrinsic.py I had to change the scenes = ["Scene01", "Scene02", "Scene18", "Scene06", "Scene20"] -> scenes = ["0001", "0002", "0018", "0006", "0020"] with the dataset I downloaded. I'm guessing I'm missing the right dataset. If so how can I get the dataset that was used?

[humenbergerm]

Hi!

I am sorry, the vkitti2 dataset (which is needed for slamantic) is not released yet. I will add this to the readme file.
It will be released soon, in the meanwhile, please use kitti.

Best,
Martin

[mshong0320]

Hi Martin,
Thank you for your reply.
Do you know when will that vkitti2 dataset be released?
Could you answer my question about using dla to generate the kitti/sequence/dla/labelIDs/, kitti/sequence/dla/labelProbabilities/ and the semantic labels for kitti?

Thank you again.

Best regards,

Chris

[humenbergerm]

Hi!

I guess beginning of December 2019.
Sorry, I did not run dla myself, I let mthz answer that.

Best,
Martin

[mshong0320]

Thank you for the reply Martin!

[mthz]

Hi,
unfortunately i cannot share the model and semantic generation we used with the DLA algorithm. but you can use any semantic labeling implementation. The input is the semantic image where I(x,y)=semantic class ID. You can use as well a colored semantic image where I(x,y)=(r,g,b)

yes you could as well use bounding boxes to gather the semantic information. But you have to take into account that 3D-Points could be within the bounding box but not on the object itself. Furthermore you have to manage overlapping boxes somehow. The semantic labeling allows to use easily background labels such as road, sidewalk, building. We directly started with a semantic labeling model and thus didn't do experiments with bounding box detectors.

Best,
Matthias

[mshong0320]

Hi Matthias,

Thank you very much for the detailed reply.

Best regards,

Chris

[humenbergerm]

Hi!

VKITTI2 is released now. I added the link to the readme file.

Best,
Martin