Argoverse waypoint improvements in junctions #1944
Conversation
|
Wondering whether it would be possible to slightly tighten the potential waypoint paths which now appear a little too loose?
|
|
The problem is that there is no road connectivity between that road and where the ego is, so there is no valid path leading back to those waypoints, though they are still part of the junction. Would it be suitable to filter out those waypoints earlier on so that they aren't so far away? |
| @@ -881,7 +881,7 @@ def closest_lanepoints( | |||
| k=maximum_count, | |||
| filter_composites=True, | |||
| ) | |||
| return [l_lps[0].lp for l_lps in linked_lanepoints] | |||
| return [l_lps.lp for l_lps in linked_lanepoints[0]] | |||
There was a problem hiding this comment.
Before
+---+---+---+
|p1 |p2 |p3 |
+---+---+---+
| x | x | x |
+---+---+---+
| | | |
+---+---+---+
After
+---+---+---+
|p1 |p2 |p3 |
+---+---+---+
| x | | |
+---+---+---+
| x | | |
+---+---+---+
Am I mistaken in thinking that now all the lps will be related to the first pose?
Maybe the method is not doing what it says it should be. It gives a maximum_count parameter but currently only gives a single lanepoint per pose. It does use maximum_count to figure out the maximum lanepoints to consider for the one lanepoint it gives per pose.
There was a problem hiding this comment.
Good catch. I didn't even realize that this takes in multiple poses. I'm a bit confused by the intention of this function, considering it seems to only ever be called with 1 pose. The original definition only took the first lanepoint of each set of 10 lanepoints for each pose. I think I may just change it to take a single pose and leave it as-is.
smarts/core/argoverse_map.py
Outdated
| return angle | ||
|
|
||
| # Filter out any waypoints that are behind the vehicle | ||
| wps = [wp for wp in path if _angle_between(pose, wp) < np.pi / 1.5] |
There was a problem hiding this comment.
This might not work because the vehicle could get turned around and lose out on seeing waypoints that are ahead of where it should go because it is turned around.
There was a problem hiding this comment.
Considering @Adaickalavan's comment on the junction waypoints I am not entirely certain what should be done.
- Filter branches by using the closest similarly near-oriented point's orientation to cull other branches
- Cut off branches which have the nearest point far away and ends oriented away.
- Leave roughly as is
There was a problem hiding this comment.
We are now filtering based on distance to first lanepoint.
I just realised that the same issue of far-away waypoints occur in the left-turn path too.
|
|
I've made some changes based on the feedback.
This should now give all possible paths, and relevant paths should not suddenly disappear. However, sometimes new paths can suddenly appear if the vehicle drives onto a new lane. I think this is fine, since an agent can determine for itself which waypoints to follow. See the video below for a demo: argoverse-junctions3.mp4 |
This addresses #1926, specifically the 3rd point raised in the issue. We now keep track of the last junction lane used for waypoints, to prevent erratic behaviour when the vehicle is on top of multiple lanes.
This also addresses #1927, by generating waypoints for all paths leading into the current junction lane:
argoverse-junctions1.mp4