problem about the prepares dataset

[YihaoX]

Dear db-net team,

Thanks for your great paper of dbnet. Sorry for sent an email with the same content to you because it is really improtent for my experiments.

After downloading and training the quick start on your GitHub, here comes 2 problem:

1. In the val data, form \dbnet-2018 \val\18\dvr_66x200 to \dbnet-2018 \val\20\dvr_66x200, the car is totally stopping, but in the
   Behaviors.csv, the speed is not zero but about around 40km/s. I don’t think that it can be such a huge measurement error.
2. I have done the quick training you provide and the val loss is smaller that train loss, could you tell me the reason for it? I also try the code of io-nvidia, it is also the same.

if you modify the dataset and the baseline, hope for reply.
thanks