Code for custom dataset

[duda1202]

Hi,

I would like to use this code for different datasets that use rostopics for images and pointclouds. I have noticed the code is made for ROS but there is no argument in the launch files for the rostopics. Is it possible to use it with other datasets?

Thank you

[doctorcolossus]

Hey @duda1202, did you ever figure this out? Any help from @ganlumomo? I am also interested in trying to use this package with my robotic platform, whose exteroceptive sensors consist of two cameras and a LiDaR sensor. The idea sounds really cool, but I am completely lost as to how it can be configured...

As @duda1202 wrote, it is not apparent where different sensor input topics could be configured. I can see that there are configuration files for methods and datasets, but the only configurable topic in any of those seems to be the (presumably output) map, entitled occupied_cells_vis_array (visited?, visual?, visible?, visualization?) by default. But what about sensor inputs?

The "toy example" uses the sim_unstructured_annotated dataset, which seems to consist only of .pcd files. But other datasets in the repository include bagfiles with topics /robot_pose (of type geometry_msgs/PoseStamped) and /selected_pc2_map (of type sensor_msgs/PointCloud2), according to rosbag info, but the paper talks about using "multimodal sensory data" and "an arbitrary number of sensors". Can other topics and modalities really be used? If so, how can that be configured?

There is also an argument in the sensor configuration yaml file called scan_num which I don't understand. There is a comment which says "how many scans to use", essentially just repeating the information which is implicit in the variable name. But what is meant by a "scan" in this context? The value in the predefined configurations is 12, and there are 12 pcd files, so is each file considered a "scan"? Or does this mean to only use the first 12 datapoints (in each file? altogether?)? The paper mentions that "the average runtime of the multilayer mapping is 1.83 s/scan for both layers", which sounds like the inference time, but that wouldn't make sense in the context of this configuration argument.

This repository could definitely use a lot more documentation to help explain how it can be used. And I would be glad to help contribute to that if anyone could help me figure out how to use it myself.