We wanted to understand the contribution of different modules in VAD. For the same, in the config VAD_base_e2e.py, we updated the loss-weight = 0.0 for all losses except the loss_plan_reg and trained the model. With this limited supervision, we obtained the following planning results at 5th Epoch and were a bit surprised.
We were wondering if these planning results (which are better than the results reported in Table 1) are expected or we are missing something in interpreting these results? Should we look more closely in the collision metrics?
Could it be possible that a pretrained BEVFormer checkpoint is getting loaded into VAD model ?
Hi @outsidercsy ,
We wanted to understand the contribution of different modules in VAD. For the same, in the config
VAD_base_e2e.py
, we updated theloss-weight = 0.0
for all losses except theloss_plan_reg
and trained the model. With this limited supervision, we obtained the following planning results at 5th Epoch and were a bit surprised.We were wondering if these planning results (which are better than the results reported in Table 1) are expected or we are missing something in interpreting these results? Should we look more closely in the collision metrics?
Could it be possible that a pretrained BEVFormer checkpoint is getting loaded into VAD model ?
Any insights/pointers would be super helpful.
Planning results :-
Thanks in advance !