Unfair comparison with previous work

[Kguo-cs]

I think your comparison is unfair. Previous work like PEC_net only generate one trajectory based on the best of 20 destinations for computing Min_ade. However, you generate 20 trajectories based on the best of 20 destinations and then choose the best ade. I evaluate your model like PEC-net. The ade reuslt is: 13.297966124938515 (rsample) and 10.666338165054066(mean).

[JoeHEZHAO]

Thanks for your interest.

PECNET is a good work and it samples 20 destinations using their well-trained CVAE, whereas this work retrieves the same amount of destinations via only doing near-nest neighbor. There is no sampling from a learned neural network on destinations, and thus sampling is focused on the other 11 trajectory points from the learned model.

Note in the paper we said, "...which decouples the goal inference from subsequent trajectory sampling and therefore reprioritizes the sampling on the overall trajectories."

[JoeHEZHAO]

I assume there is no further question. Feel free to reopen it.