Fast Kernel Scene Flow

Xueqian Li and Simon Lucey.

arXiv link: https://arxiv.org/abs/2403.05896

This is the code implementation of Fast Kernel Scene Flow, a fast per-point embedding-based kernel method for scene flow estimation.

---

With input point cloud $\mathcal{S}_1$, $\mathcal{S}_2$, we could use different approaches to extract per-point features. For example: (a) Raw points, which treat original points as point features; (b) PE, which uses an RFF-based positional encoding to embed the input point to high-frequency features; (c) PEAT, which extracts learned point features using positional encoding and self-attention. A kernel function $\mathcal{K}$ is then employed to compute the similarity matrix between these two inputs based on their point features. Finally, a linear coefficient vector $\mathbf{\alpha}$ is optimized per sample to predict the final flow. Our model is compact and fast, with only $\mathbf{\alpha}$ being the learnable parameter during runtime.

---

Prerequisites

This code is based on PyTorch implementation, and tested on PyTorch=1.12.0, Python=3.9.15 with CUDA 11.6.

A simple installation is bash ./install.sh.

For a detailed installation guide, please go to requirements.yml.

Dataset

We directly use the datasets provided by FastNSF. You may download datasets used in the paper from these links:

• Argoverse (390MB)

• Waymo Open (476MB)

---

Running the code

• Argoverse scene flow dataset

  python kernel_flow.py --device cuda:0 --dataset_name argoverse --data_path <dataset_dir> --iters 1000 --earlystopping --early_patience 10 --early_min_delta 0.001 --kernel_grid --grid_factor 0.2 --model pe --weight_decay 0. --use_dt_loss --dt_grid_factor 10. --use_all_points --alpha_init_method same_as_linear --alpha_init_scaling 1. --reg_name l1 --reg_scaling 5. --epsilon 1e-7 --pe_type RFF --pe_dim 256 --pe_sigma 0.01 --log_sigma 10. --alpha_lr 0.008
  

• Waymo Open scene flow dataset

  python kernel_flow.py --device cuda:0 --dataset_name waymo --data_path <dataset_dir> --iters 1000 --earlystopping --early_patience 10 --early_min_delta 0.001 --kernel_grid --grid_factor 0.2 --model pe --weight_decay 0. --use_dt_loss --dt_grid_factor 10. --use_all_points --alpha_init_method same_as_linear --alpha_init_scaling 1. --reg_name l1 --reg_scaling 5. --epsilon 1e-7 --pe_type RFF --pe_dim 256 --pe_sigma 0.01 --log_sigma 10. --alpha_lr 0.008
  

---

Acknowledgement

Neural Scene Flow Prior

Fast Neural Scene Flow

FastGeodis: Fast Generalised Geodesic Distance Transform

Contributing

If you find the project useful for your research, you may cite,

@article{li2024fast,
  title={Fast Kernel Scene Flow},
  author={Li, Xueqian and Lucey, Simon},
  journal={arXiv preprint arXiv:2403.05896},
  year={2024}
}