Distance to closest semantic wall

[eriksandstroem]

❓ Questions and Help

Hi,
Thanks for providing Habitat! Great work!

I am building an agent that walks around in the scene by traversing randomly navigable points and collecting sensor data from the Replica dataset.

I have successfully achieved the above, but I would like to have better control when sampling the random points. I do not want them to be too close to a wall because this will make the sensor data less reliable. Being close to e.g. a chair is, however, not a problem. Currently, I have set a threshold as follows:

traverseList = []
randPt = sim.pathfinder.get_random_navigable_point()
dist_to_obs = sim.pathfinder.distance_to_closest_obstacle(randPt, max_search_radius)
if dist_to_obs >= max_search_radius:
    traverseList.append(randPt)

This means that I will add the point only if its distance to any obstacle is at least max_search_radius. This means that I will also stay away from furniture, which I don't want.

My question: Given a surface point (e.g. coming from the output of the function "closest_obstacle_surface_point", is there any way to find it's semantic class such that I can only focus on walls?

Thanks a lot

[erikwijmans]

My question: Given a surface point (e.g. coming from the output of the function "closest_obstacle_surface_point", is there any way to find it's semantic class such that I can only focus on walls?

In theory yes, you can use the bounding boxes in the semantic annotations to approximate this. However, in practice, this still won't give you what you are looking for as the closet obstacle search will simply find the closet obstacle, regardless of whether or not it is furniture or a wall, so you may still end up near walls.

Perhaps a better way to do this is to iterate over all the wall objects in the scene (https://aihabitat.org/docs/habitat-sim/habitat_sim.scene.SemanticScene.html#objects from here: https://aihabitat.org/docs/habitat-sim/habitat_sim.simulator.Simulator.html#semantic_scene) and then check to see if the distance to the bounding box is greater than some threshold (https://aihabitat.org/docs/habitat-sim/habitat_sim.geo.OBB.html#distance). I believe the walls in replica are annotated in a way that they can be well approximated by a bounding box, so this should work.

[dhruvbatra]

I am building an agent that walks around in the scene by traversing randomly navigable points and collecting sensor data from the Replica dataset.

Sounds interesting.

One pointer -- @mpiseno is building native "data/sensor acquisition" capability in Habitat-Sim, i.e. the ability to extract images/semantics from various viewpoints across 3D scenes. You can follow his progress here: #518

[mathfac]

Closing as the question was answered.