search

yuanjun-gao / ICCV2023_SLAM_Challenge

73 stars 2 forks source link

General Feedback #24

Closed davidskdds closed 9 months ago

davidskdds commented 11 months ago

Hello, I am PhD student in the SLAM area and wanted to participate in your SLAM challenge. After attending in the Hilti SLAM challenges 2022 and 2023 with my pipeline I also took a look at your datasets. Unfortunately, I noticed a few difficulties that, in my case, reduced the motivation to participate:

  1. the sequences from the TartanAir dataset lack timestamps per point. Only in another Github issue it was communicated that the scans were generated statically in each case. However, the static generation of the scans does not fit the reality and prevents the easy application of common LiDAR inertial SLAM approaches.

  2. the sequences from the TartanAir dataset are multiple times provided with a different noise identifier. Here it is unclear which Rosbags are actually relevant.

  3. only velodyne scan messages are included in the Rosbags from the SubT-MRS dataset. This is a manufacturer-specific format. Here it would be easier if the data had already been converted into PointCloud2 messages in advance, as this format is more common.

All in all, due to the points mentioned above, participation entails some additional development effort, which in my case reduces the willingness to participate.

I don't want to criticize you, but to share my perception in a productive way, in order to possibly reach more participants in future SLAM Challenges.

Kind regards David Skuddis

shibowing commented 11 months ago

Hi @davidkdds,

Thank you so much for the feedback! We appreciate your kind advice.

We will provide the solutions in the next 36 hours.

Best,

Shibo

yuanjun-gao commented 11 months ago

Instructions for converting velodyne_msgs/VelodyneScan to sensor_msgs/PointCloud2 are updated on README.md

shibowing commented 11 months ago

Hi @davidskdds ,

Thank you for your feedback. We have made some quick fixes to address your questions and hope that these improvements enhance your overall participation experience

  1. I have reached out to Wenshan regarding the Tartan Air datasets and the timestamp issue. As the data is collected in simulation environments, each data point is timestamped simultaneously. However, I acknowledge your point, and we will address this concern by providing new datasets that aim to better simulate real-world lidar scans.

  2. That confused datasets are removed in the google drive. Thanks for feedback

  3. Since our datasets containts multi-modalities, the PointCloud2 messages will make rosbag very large and takes a lot time to download. Therefore, we added intructions to convert the velodyne_msgs/VelodyneScan to sensor_msgs/PointCloud2. This is just simply launch a rosnode to subscribe the VelodyneScan and publish PointCloud2 message. I think it is easy to install and hope it can make your life easier.

Instructions for converting velodyne_msgs/VelodyneScan to sensor_msgs/PointCloud2 are updated on README.md

Thanks again for your suggestions. Feel free to let me know if you have other suggestions! The datasets is still improving and maintaining for long-term!