Skip to content

Clarification on MTR checkpoints/configs and Kalman difficulty bins for Table 7 #73

Description

@OldSaltyFish-Handsome

Hi UniTraj authors,

Thank you for releasing UniTraj. I am currently studying the MTR experiments in the paper and trying to understand/reproduce the analysis in Table 7 of arXiv v3, where MTR trained on nuScenes and MTR trained on all datasets are compared on the nuScenes validation set by Kalman difficulty.

The reported Brier-minFDE values are:

MTR (nuScenes): 2.73 / 4.52 / 4.25
MTR (All): 2.23 / 2.97 / 4.20

for Easy / Medium / Hard samples.

I have a few questions about the provenance of this experiment:

Are the checkpoints corresponding to MTR (nuScenes) and MTR (All) from Table 7 still available?
I understand from Issue #17 that pretrained checkpoints involving Waymo/WOMD may not be publicly releasable because of licensing restrictions.
If the all-data checkpoint cannot be shared, would it be possible to provide the nuScenes-only checkpoint, or confirm whether it is also unavailable?
If the checkpoints cannot be shared, would it be possible to provide or confirm the resolved training/evaluation configuration used for the two Table 7 runs, especially:
datasets used in the “All” condition,
training/validation split,
training epochs / checkpoint-selection rule,
relevant preprocessing / ScenarioNet settings,
evaluation configuration.

I noticed a discrepancy between the Kalman-difficulty bins reported in Table 7 and the current public code:

Table 7: Easy [0,30), Medium [30,50), Hard [50,100)
current code: Easy [0,30), Medium [30,60), Hard [60,+∞)

Could you please confirm which bin definition was actually used to produce Table 7?

If the original checkpoints/configs are no longer available, even a brief confirmation of the actual Table 7 setup and difficulty-bin definition would be extremely helpful for avoiding an incorrect reproduction.

My goal is not to redistribute any restricted checkpoint, but to understand the exact experimental setup and reproduce the failure analysis as faithfully as possible.

Thank you very much for your time and for making UniTraj publicly available.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions