Dear @MasterIzumi
I noticed that your loss function calculated all agents in a scenario but not only ego agent. On AV2 single agent motion forecasting leaderboard, SIMPL got 1.43 minFDE6 metrics which seems great. I wonder if it could get more accuracy on single agent motion forecasting leaderboard when involving all agents into loss function but not only ego agent?Other SOTA work like QCNet utilizing all agents' ground truth trajectory in loss function as well.
In fact, I studied some previous work like Autobot (2022).The original code of Autobot didn't take RPE into encoder.After i add rpe module into Autobot, it got only 1.7 minFDE6 on AV2 validation dataset.I wonder if it's because i only use ego agent into loss function?Have you ever explore this?
I'd appreciate if you could reply!Thank you so much.
Dear @MasterIzumi I noticed that your loss function calculated all agents in a scenario but not only ego agent. On AV2 single agent motion forecasting leaderboard, SIMPL got 1.43 minFDE6 metrics which seems great. I wonder if it could get more accuracy on single agent motion forecasting leaderboard when involving all agents into loss function but not only ego agent?Other SOTA work like QCNet utilizing all agents' ground truth trajectory in loss function as well. In fact, I studied some previous work like Autobot (2022).The original code of Autobot didn't take RPE into encoder.After i add rpe module into Autobot, it got only 1.7 minFDE6 on AV2 validation dataset.I wonder if it's because i only use ego agent into loss function?Have you ever explore this? I'd appreciate if you could reply!Thank you so much.