Closed zhepanqi closed 1 month ago
Thank you so much for your work and confirmation!! In terms of DenseTNT training, yes we are still using the same files as in HiVT. For specific training procedure, I might need to consult with my collaborator and let you know later.
Hello, It's me again. We have just completed the training of your HiVT model and successfully achieved the results. We are now working on the training of DenseTNT. Could you please confirm if you are still using the files generated by adaptor.py from the Nuscenes dataset (the same as for HiVT)? Additionally, why does your train_and_eval.sh script include 'goals_2D' in the other_params? Shouldn't this be a parameter for the Argoverse dataset? It seems that training with this setup may not allow for proper evaluation in the end.
Hi Guantian, thank you for the comment. "goals_2D" refers to the goal-based strategy densetnt utilizes to predict the trajectories. DenseTNT estimates the goals (end positions) of agents' future trajectory and then generates the complete trajectories based on the goals. There do be some redundant arguments in the script but none of them should affect anything. I will also clean the repo up a little bit shortly.
Hello, It's me again. We have just completed the training of your HiVT model and successfully achieved the results. We are now working on the training of DenseTNT. Could you please confirm if you are still using the files generated by adaptor.py from the Nuscenes dataset (the same as for HiVT)? Additionally, why does your train_and_eval.sh script include 'goals_2D' in the other_params? Shouldn't this be a parameter for the Argoverse dataset? It seems that training with this setup may not allow for proper evaluation in the end.