Scene Flow Zoo

Scene Flow Zoo is an open source scene flow benchmark and model zoo. Based on the ZeroFlow codebase, it uses the BucketedSceneFlowEval to load datasets and perform evaluation.

Currently, the Zoo supports the following datasets:

• Argoverse 2
• Waymo Open

The Zoo supports the following methods:

• Feed-forward
  • FastFlow3D / FastFlow3D XL
  • ZeroFlow and ZeroFlow XL
  • DeFlow
• Test time optimization
  • Neural Scene Flow Prior (NSFP)
  • Fast NSF
  • Liu et al. 2024
  • EulerFlow

If you use this codebase, please cite the following paper:

@article{vedder2024zeroflow,
    author    = {Kyle Vedder and Neehar Peri and Nathaniel Chodosh and Ishan Khatri and Eric Eaton and Dinesh Jayaraman and Yang Liu Deva Ramanan and James Hays},
    title     = {{ZeroFlow: Fast Zero Label Scene Flow via Distillation}},
    journal   = {International Conference on Learning Representations (ICLR)},
    year      = {2024},
}

If you use the evaluation results, please cite the following paper:

@misc{khatri2024trackflow,
    author = {Ishan Khatri and Kyle Vedder and Neehar Peri and Deva Ramanan and James Hays},
    title = {I Can't Believe It's Not Scene Flow!},
    journal = {arXiv},
    eprint = {2403.04739},
    year = {2024},
    pdf = {https://arxiv.org/abs/2403.04739}
}

If you use any of the methods in the Zoo, please cite the appropriate paper for that method.

Pre-requisites / Getting Started

Read the Getting Started doc for detailed instructions to setup the datasets and use the prepared docker environments.

Pretrained weights

Trained weights are available for the following methods:

• ZeroFlow and FastFlow3D
• DeFlow

Visualizing results

The visualization/visualize_flow.py script can visualize the ground truth flow and the predicted flow for various methods. Note that the visualizer requires the ability to start an X window; the ./launch.sh script on a headed machine will do this for you.

Training a model

Inside the main container (./launch.sh), run the train_pl.py with a path to a config (inside configs/) and optionally specify any number of GPUs (defaults to all GPUs on the system).

python train_pl.py <my config path> --gpus <num gpus>

The script will start by verifying the val dataloader works, and then launch the train job.

Note that config files specify the batch size per GPU, so the effective batch size will be batch_size * num_gpus. In order to replicate our results, you must use the effective batch size of 64 for the normal sized FastFlow3D-style model and an effective batch size of 12 for the XL model. Our configs are setup to run on 4 x A6000s for the normal model and 6 x A6000s for the XL model. If your system differs, set the accumulate_grad_batches parameter in the config to accumulate gradients over multiple batches to reach the same size effective batch.

Testing a model

Inside the main (./launch.sh), run the test_pl.py with a path to a config (inside configs/), a path to a checkpoint, and the number of GPUs (defaults to a single GPU).

python test_pl.py <my config path> <my checkpoint path> --gpus <num gpus>

Submitting to the AV2 2024 Scene Flow competition

1. Dump the outputs of the model for the test split
   • Run test_pl.py with a dumper config that has the save_output_folder set to the desired output folder, and the test set as the val_split (e.g. configs/fastflow3d/argo/bucketed_nsfp_distillation_3x_test_dumper)
   • The dumper must be run with a single GPU (the default), as some batch entries may be skipped with multi-GPU inference.
2. Build the competition submission the output with python av2_scene_flow_competition_submit.py <path/to/dumped/output/folder/>
   • This will create a zip file as a sibling in the filesystem to the output folder, named after it.
3. Submit the zip to the competition website.
   • EvalAI's CLI must be used, as the zip file exceeds the limit for web uploads.