SHARP: Short-Window Streaming for Accurate and Robust Prediction in Motion Forecasting
Alexander Prutsch, Christian Fruhwirth-Reisinger, David Schinagl, Horst Possegger
Graz University of Technology
CVPR 2026

This repository provides full data preprocessing, training and inference support for the nuScenes, Argoverse 1 (AV1), and Argoverse 2 (AV2) datasets.
It also includes pretrained checkpoints for AV2 single- and multi-agent settings, and visualization tools for AV2.

Getting Started

Create and Activate Virtual Environment

conda create -n sharp python=3.11
conda activate sharp

Install PyTorch

We tested our implementation with torch 2.8.0 and CUDA 12.8

Install PyTorch e.g.

pip install torch==2.8.0 torchvision --index-url https://download.pytorch.org/whl/cu128

Install Dependencies

pip install -r ./requirements.txt

Dataset Setup

Download the Argoverse 2 Motion Forecasting Dataset

The expected structure of the AV2 data should be:

data_root
    ├── train
    │   ├── 0000b0f9-99f9-4a1f-a231-5be9e4c523f7
    │   ├── 0000b6ab-e100-4f6b-aee8-b520b57c0530
    │   ├── ...
    ├── val
    │   ├── 00010486-9a07-48ae-b493-cf4545855937
    │   ├── 00062a32-8d6d-4449-9948-6fedac67bfcd
    │   ├── ...
    ├── test
    │   ├── 0000b329-f890-4c2b-93f2-7e2413d4ca5b
    │   ├── 0008c251-e9b0-4708-b762-b15cb6effc27
    │   ├── ...

Download the Argoverse 1 Motion Forecasting Dataset

nuScenes

Please follow the official guidelines.

Data Preprocessing

Preprocess the Argoverse 2 dataset by executing

python preprocess_av.py --data_root=/path/to/av2_data_root -p

Preprocess the Argoverse 1 dataset by executing

python preprocess_av.py --data_root=/path/to/av1_data_root -p --av1

For nuScenes please run the standalone nus_extractor.py script in the src/datamodules folder.

Training on Single-Agent Benchmarks

Train SHARP model on single-agent data using

python train.py datamodule.pl_module.data_root=/path/to/data_root/sharp_processed/

Select the data root for the corresponding dataset.

Evaluation on Single-Agent Benchmarks

Evaluate SHARP model using

python eval.py datamodule.pl_module.data_root=/path/to/data_root/sharp_processed/ checkpoint=/path/to/checkpoint.ckpt

Select the data root for the corresponding dataset.

AV2 Single-Agent Checkpoint provided in the repository.

python eval.py datamodule.pl_module.data_root=/path/to/av2_data_root/sharp_processed/ checkpoint=exps/av2_single_agent/checkpoints/av2_sa.ckpt

Expected results:

AV2 Single-Agent Visualization

Visualize the prediction results using

python visualize_av2_sa.py

Please update the data_root, chkpt_dir, and av2_raw_data_dir variable in the script.

Training and Evaluation on AV2 Multi-Agent Benchmark

1. Update config_name to "config_ma" in train.py
2. Initialize single-agent model with your checkpoint (see TODO in sharp.py)
3. Train SHARP model with multi-agent consistency module using

python train.py datamodule.pl_module.data_root=/path/to/av2_data_root/sharp_processed/

Evaluation on AV2 Multi-Agent Benchmark

1. Update config_name to "config_ma" in eval.py
2. Evaluate SHARP model using

python eval.py datamodule.pl_module.data_root=/path/to/av2_data_root/sharp_processed/ checkpoint=/path/to/checkpoint.ckpt

AV2 Multi-Agent Checkpoint provided in the repository.

python eval.py datamodule.pl_module.data_root=/path/to/av2_data_root/sharp_processed/ checkpoint=exps/av2_multi_agent/checkpoints/av2_ma.ckpt

Expected results:

This checkpoint differs from our final challenge submission which is trained on both train and validation set.

Bibtex

@inproceedings{prutsch2026sharp,
    title={{SHARP: Short-Window Streaming for Accurate and Robust Prediction in Motion Forecasting}},
    author={Prutsch, Alexander and Fruhwirth-Reisinger, Christian and Schinagl, David and Possegger, Horst},
    booktitle={In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    year={2026}
}

Acknowledgements

This repository is based on RealMotion and integrates code from Forecast-MAE, DeMo, and EMP. We thank them for their work!