Closed huangju91 closed 4 years ago
Hi @huangju91, the bag was generated with a simulator developed with our colleagues at LincolnLabs. We are planning to release it for a workshop competition at ICRA, I’ll update this issue with the official release. Thanks!
Hi @huangju91, the bag was generated with a simulator developed with our colleagues at LincolnLabs. We are planning to release it for a workshop competition at ICRA, I’ll update this issue with the official release. Thanks!
I found the corresponding description in the Kimera paper that you utilized a simulator to generate the ground truth and 2-D semantic images so as to evaluate the Kimera-Semantics performance. Meanwhile you recommended to use deep learning neural networks to generate 2d semantic labels. While I am confused that when the stereo-dense-reconstruction-node was turned on which you said to utilize dense stereo to obtain 3d point clouds in the paper, I could not found where the depth information was calculated from the raw stereo image inputs. I tried to search in the Kimera-VIO module to find the dense stereo depth calculation processes but I got some trouble in the installation. So would you plz help me figure out this issue? Thanks much for the aswsome work you shared.
@XanxusCrypto checkout the fag run_stereo_dense
in here: https://github.com/MIT-SPARK/Kimera-Semantics/blob/6637d8b54c1fc17903f507ff0f536397eb107d3d/kimera_semantics_ros/launch/kimera_semantics.launch#L38
If you turn on this flag in the launch file (and you have a pair of stereo images, Kimera-Semantics will make use of depth coming from the stereo reconstruction algorithm, but you could technically use any kind of point-cloud even a LIDAR based one like in here (where I used Semantic-Kitti):
@ToniRV Thanks much for answering my questions. So the stereo reconstruction algorithm is constructed in Volblox? I just read the Volblox paper which utilized stereo or RGB-D camera, but no more useful information was found further in their launch file of EuRoC dataset. While in their newer version of Volblox++, only RGB-D camera is utilized. Bcs I would like to run visual-inertial SLAM with stereo camrea to output poses and points, I would like to understand how and where was the dense stereo reconstruction algorithm constructed. Thank you again for sharing your code.
The stereo reconstruction is done using ROS image_proc package, in particular using stereo_image_proc which under the hood uses StereoBM algorithm implemented in OpenCV.
The stereo reconstruction is done using ROS image_proc package, in particular using stereo_image_proc which under the hood uses StereoBM algorithm implemented in OpenCV.
Thanks very much for the elaborative answers. I have found the dense stereo reconstruction algorithm utilized. It helped me a lot!
Closing on this issue: the simulator is the same that we will be using in our ICRA2020 workshop this May: https://mit-spark.github.io/PAL-ICRA2020/
Hi @ToniRV,
The simulator you mentioned in the last comment, was it the same one used for the goseek-challenge?
Also, is there a recording or any other material related to that ICRA2020 workshop available? It would be a great resource.
I found the recording of the workshop. Posting the link here: Metric-Semantic SLAM with Kimera: A Hands On Tutorial
This work is amazing. Thanks for making it open-source!
Hi,Thanks for sharing this awesome work, I wonder how you create the kinera_semantics_demo.bag, is it made from another module like Kimera-VIO ?