Junru Gu
Junru Gu
We treat sampling on the centerline as sparse sampling. Dense sampling is performed at https://github.com/Tsinghua-MARS-Lab/DenseTNT/blob/a0e3b8a51aecf9f9046db4fb72e2793684c96e69/src/modeling/decoder.py#L143
`predict_traj` here is only used to calculate loss for trajectory completion. During training, we don't need to generate top K goals. During evaluation, we get top K goals at https://github.com/Tsinghua-MARS-Lab/DenseTNT/blob/a0e3b8a51aecf9f9046db4fb72e2793684c96e69/src/modeling/decoder.py#L263
We haven't tested it on Argoverse 2 yet. If it goes well, we will update results on Argoverse 2 in August.
We are tuning DenseTNT on Argoverse 2. We add the training code of Argoverse 2 for reference: https://github.com/Tsinghua-MARS-Lab/DenseTNT/tree/argoverse2
> > We are tuning DenseTNT on Argoverse 2. We add the training code of Argoverse 2 for reference: https://github.com/Tsinghua-MARS-Lab/DenseTNT/tree/argoverse2 > > Hi, do you have plans for updating the...
> > > > We are tuning DenseTNT on Argoverse 2. We add the training code of Argoverse 2 for reference: https://github.com/Tsinghua-MARS-Lab/DenseTNT/tree/argoverse2 > > > > > > > >...
Instead of using the waymo dataloader, we use the parsing function provided in the waymo tutorial to parse raw data into a dictionary, and then convert the tf tensors in...
We use the MIT license for this project.
We did not release pretrained model since the best performance can be easily reproduced by running our training command. The video demos are created by Bokeh.
Thanks for your suggestion. We have added the license file.