autoware_perception_evaluation icon indicating copy to clipboard operation
autoware_perception_evaluation copied to clipboard

Published 20 hours ago • verified publisher icon tier4

feat: update matching policy

Open ktro2828 opened this issue 1 year ago • 0 comments

Category

  • [x] Perception

What

Update the matching policy as follows:

  1. Matching the nearest estimated objects and GTs which have the same label primary.
  2. Matching the nearest estimated objects and GTs regardless of their label.

For UseCase datasets, it contains unintended FPs. Therefore in order to prevent them, we should specify max_matchable_radii: <float | list[float]>. max_matchable_radii can be specified for each specific label.

Before PR

object_matching1

After PR

BasicObjectsMatching_v2 0

NOTE

As shown in the above fig (After PR), this feature will increase FPs which are not matched any GTs. In order to prevent this problem, this PR also add a new function, named filter_unintended_object_results(...). This function filters out these FP results.

FilterUnintendedResults

As a future work, it is helpful if we can only filter out unintended results, which appear in UseCase data, except of ghost objects.

Type of change

  • [x] New feature

Test performed

  • [ ] test.sensing_lsim
  • [ ] test.perception_lsim

Reference

Notes for reviewer

ktro2828 avatar May 10 '24 09:05 ktro2828