search

ynw2021 / FEND

6 stars 0 forks source link

Discrepency between the cited traj+EWTA results and the original ones #3

Open ArielleZhang opened 11 months ago

ArielleZhang commented 11 months ago

Hi,

I noticed that in Table 2 in the FEND paper, it is said that the results are obtained from the pretrained model from the "On Exposing the Challenging Long Tail in Future Prediction of Traffic Actors", for example on nuScenes, both evaluated on Vehicles, the top1% for trajectron++ + EWTA + contrastive reported in FEND is 1.28/2.85 (minADE/minFDE) but in the "On Exposing..." paper the result is 0.54/0.85 (minADE/minFDE), could you elaborate on where you get the results?

Thanks,

Arielle

ArielleZhang commented 10 months ago

I realized you are using different metrics (FDEs of the baseline Trajectron++ + EWTA) to rank the scene difficulties, apologies for the misunderstanding! However, I tried using the mean of the best-of-20 FDEs, minFDE and maxFDE, I still couldn't reproduce the results, could you elaborate a bit more on the details of how you rank the challenging scenes for the experiment results? Thanks in advance!

ArielleZhang commented 10 months ago

Just to follow up, I was working on nuScenes, not pedestrians dataset. Details about the nuScenes dataset would be helpful!