Towards Metric-Agnostic Trajectory Forecasting

ECCV 2026

arXiv | Project Page | BibTeX

Markus Knoche¹, Daan de Geus², Bastian Leibe¹

¹ RWTH Aachen University ² Eindhoven University of Technology

Note

This repository contains code for our model DONUT-NLL. To apply and evaluate model samples using our TraDiE policies, check out this repository.

Installation

Clone repository:

git clone https://github.com/MKnoche/DONUT-NLL.git
cd DONUT-NLL

Install uv.

Data Preprocessing

Download the raw Waymo data into {args.data_root}/waymo/raw/<split>, where <split> is training, validation, or testing.

Next, preprocess the data. This will take some time.

cd preprocessing
uv run preprocess_waymo.py --data_root <data_root>

Training

Adjust the root paths in train_donut_nll.py according to your setup.

Distributed training is supported via the devices and nodes parameters. Gradient accumulation makes sure that the effective batch size is always acc_batch_size, as long as batch_size * devices * nodes <= acc_batch_size.

Training for 30 epochs on 6 NVIDIA H100 GPUs with a batch size of 8 per GPU takes about 2.7 days.

uv run train_donut_nll.py

Evaluation

First, generate and store samples from the model. The newest checkpoint is automatically loaded from {args.ckpt_root}/{args.model_name}/.

uv run sample_donut_nll.py <model_name>

Next, follow the steps in the TraDiE-policy repo.

Model Checkpoint

DONUT-NLL (Generalized Gaussian + Step-NLL)

Download and extract the checkpoint in {args.ckpt_root}/donut-nll/epoch=29-step=228300.ckpt.

Citation

If you use our work in your research, please use the following BibTeX entry.

@inproceedings{knoche2026tradie,
  title     = {{Towards Metric-Agnostic Trajectory Forecasting}},
  author    = {Knoche, Markus and de Geus, Daan and Leibe, Bastian},
  booktitle = {ECCV},
  year      = {2026}
}

Acknowledgements

This project builds upon code from DONUT and QCNet (Apache-2.0 License).