Argoverse 2.0 Scene Flow Dataset

A set of scripts for working with sceneflow pseudolabels for Argoverse 2.0 Sensor

Installing

• Download and set up the Argoverse 2.0 Sensor dataset
• Clone this repository
• In this directory run python create.py --argo_dir <path/to/argo/sensor/> --output_dir <output/directory>

Evaluating

• Output your predictions as single Nx3 numpy arrays saved to a directory with the format <log_id>_.npy
• Run python eval.py --git-dir <path/to/sceneflow/files> --pred-dir <path/to/predictions>. This produces a file results.parquet which contains a pandas Dataframe with all the results. By default it breaks down the error metrics into Dynamic/Static and Foreground/Background classes, you can change the classes with the argument --breakdown.