Open ChengkaiYang opened 2 days ago
How about using 'interested agents' into training in AV2 dataset? Could this increase accuracy?
Hi,
We are using the 'focal agent' in AV2 to train the model.
UniTraj only reports vehicles' bfde in the paper, this might cause the performance discrepancy
Hi,
We are using the 'focal agent' in AV2 to train the model.
UniTraj only reports vehicles' bfde in the paper, this might cause the performance discrepancy
Thanks for your immediate nice reply! Moreover, I'd like to know what's the meaning of 'interested agent'?I noticed that all 'interested agents' in AV2 have full horizons(110, 50 for past and 60 for future).In base_dataset.py, it got preprocessed scenario in mdsn.How did we get 'interested agents'?Here is the defination in AV2 dataset. Does 'interested agents' means 'Scored Track' here?
interested agent == FOCAL_TRACK
Dear @Alan-LanFeng Glad to communication this excellent work with you again! In Unitraj, you provided three classical framework Autobot, MTR and WayFormer. I've reproduced Autobot's results on AV2 dataset, which got almost same brier-FDE(2.5) to result in your paper. However, on AV2 leaderboard, it didn't show competitive among models. After taken serious consideration, I've found all models' loss function in Unitraj were calculated on only one ego agent but not among all agents in scenario. In fact, some previous SOTA methods on AV2 single agent motion forecasting leaderboard like QCNet utilizing all agents in loss function but not only the ego one. Do you think this limited accuracy on single agent's motion forecasting or not?