Hi Christoph, will it work with a camera that is not a depth camera? The industrial camera we will be using is not a depth camera and we have been trying to find the 3D bounding annotation software for it. The search led us to your video. Any help is appreciated
Hello Gurpal, it is not dependent on the camera but only on the data it produces. So if you get a point cloud in the end in one of the supported formats (*.ply, *.pcd, *.bin, ...) it will work. The method how the point cloud is created does not really matter. Hope that helps.
labelCloud is now available on the PyPI (pypi.org/project/labelCloud/) and can be installed via pip: 1. pip install labelCloud 2. labelCloud Thus making the setup even more easier. Make sure you have a folder named "pointclouds" in the same directory or adapt the settings to your needs.
Hello, great work ! I'm planning to label pointcloud for 3D object detection. I've tried labelCloud, it seems that my pointclouds are smaller than the example therefore it works well with "big objects" but when it comes to small objects, the precision is low. It would be great if I could manually enter the dimension shown on the right box to adjust it exactly how it should be ! Or a feature that allows me to change the relative scale between the pointcloud and the steps of the boundingbox size ! Thanks !
Hi Ilias, glad to hear you are using labelCloud. You're right about the scaling. I provided some options in the config file to adjust the manipulation steps for translation, rotation, … . They are described in the documentation (github.com/ch-sa/labelCloud/blob/main/docs/documentation.md). Just change the relevant parameters (like STD_SCALING) in the config.ini. For the manual entering of parameters, I think it is a reasonable feature. Could you please open a feature request in GitHub and I might have the time to implement that feature soon. ;)
Hi Samah, the automatic labeling (=3D object detection) is the goal of this tool. But for fully automatic labeling you need to train a neural network first. And to train a network you need training data. labelCloud supports you in creating this training data. Yet it still needs your manual input for this.
@@e1nste1in first , Thanks a lot for your reply , I created trained data using labelcloud and I have pcd file so what is the next step I need to follow
Hi Somu, yes, labelCloud allows you to create a labeled training data set with 3D bounding boxes. You can then use these 3D labels to train a 3D object detection framework (like VoteNet) to detect objects in unseen 3D point clouds with their position and dimension (some frameworks can also predict the rotations).
@@christophsager7533 That sounds fantastic. Through this will void the use of openCV. Another question. Lets say if i train to detect a cube, will it detect and dimension different sizes of cube?
@@somusundram1823 Yes, it should. Though this depends on the implementation of the object detection framework. In general, the training data that you generate with labelCloud contains the specific dimensions (length, width & height) of every bounding box (i.e. object) that you labeled.
@@christophsager7533 Thank you. The reason behind my question is that i have Intel Realsense depth camera. I plan to use PCL to detect boxes and find it H x W x L with or without texture. I believe such approach gives reliability rather than RGBD. What was your opinion on this? You have stated that open for further testing / research. I plan to try yours if possible
@@christophsager7533 And also,is there any tutorial on how to export the trained data and be used in real world with depth cam or LiDAR? It would be really2 helpful.Thank you
Hi Christoph, will it work with a camera that is not a depth camera? The industrial camera we will be using is not a depth camera and we have been trying to find the 3D bounding annotation software for it. The search led us to your video. Any help is appreciated
Hello Gurpal,
it is not dependent on the camera but only on the data it produces. So if you get a point cloud in the end in one of the supported formats (*.ply, *.pcd, *.bin, ...) it will work.
The method how the point cloud is created does not really matter.
Hope that helps.
labelCloud is now available on the PyPI (pypi.org/project/labelCloud/) and can be installed via pip:
1. pip install labelCloud
2. labelCloud
Thus making the setup even more easier. Make sure you have a folder named "pointclouds" in the same directory or adapt the settings to your needs.
Hello, great work !
I'm planning to label pointcloud for 3D object detection. I've tried labelCloud, it seems that my pointclouds are smaller than the example therefore it works well with "big objects" but when it comes to small objects, the precision is low. It would be great if I could manually enter the dimension shown on the right box to adjust it exactly how it should be ! Or a feature that allows me to change the relative scale between the pointcloud and the steps of the boundingbox size !
Thanks !
Hi Ilias,
glad to hear you are using labelCloud. You're right about the scaling. I provided some options in the config file to adjust the manipulation steps for translation, rotation, … . They are described in the documentation (github.com/ch-sa/labelCloud/blob/main/docs/documentation.md). Just change the relevant parameters (like STD_SCALING) in the config.ini.
For the manual entering of parameters, I think it is a reasonable feature. Could you please open a feature request in GitHub and I might have the time to implement that feature soon. ;)
@@christophsager7533 Indeed, I've found the scaling option, thank you!
I will post the feature request though.
Hi Christoph，Is there any labeling software that can label point clouds and pictures at the same time and generate labeling files like Kitti？
Hi Christoph , I'm searching for app to automatic label the lidar 3d pcd file , can you guide me plz ?
Hi Samah,
the automatic labeling (=3D object detection) is the goal of this tool. But for fully automatic labeling you need to train a neural network first. And to train a network you need training data.
labelCloud supports you in creating this training data. Yet it still needs your manual input for this.
@@e1nste1in first , Thanks a lot for your reply , I created trained data using labelcloud and I have pcd file so what is the next step I need to follow
nice work! I plan to develop a autolabel tools base this awesome tool!
Hello Christoph. Is it possible to annotate dynamic objects? Does it have tracking functionality? thanks!
Coincidence! we are looking for the same tool for tracking objects in point clouds. Have you found such a desired tool？
Hi Chris,
Question. You are saying with depth image or PCL i can train to detect shape(s) and find its dimensions?
Hi Somu,
yes, labelCloud allows you to create a labeled training data set with 3D bounding boxes. You can then use these 3D labels to train a 3D object detection framework (like VoteNet) to detect objects in unseen 3D point clouds with their position and dimension (some frameworks can also predict the rotations).
@@christophsager7533 That sounds fantastic. Through this will void the use of openCV.
Another question. Lets say if i train to detect a cube, will it detect and dimension different sizes of cube?
@@somusundram1823 Yes, it should. Though this depends on the implementation of the object detection framework.
In general, the training data that you generate with labelCloud contains the specific dimensions (length, width & height) of every bounding box (i.e. object) that you labeled.
@@christophsager7533 Thank you. The reason behind my question is that i have Intel Realsense depth camera.
I plan to use PCL to detect boxes and find it H x W x L with or without texture. I believe such approach gives reliability rather than RGBD.
What was your opinion on this? You have stated that open for further testing / research. I plan to try yours if possible
@@christophsager7533 And also,is there any tutorial on how to export the trained data and be used in real world with depth cam or LiDAR? It would be really2 helpful.Thank you