EzFlow

A modular PyTorch library for optical flow estimation using neural networks

Documentation | Tutorials

Installation

From source (recommended)

git clone https://github.com/neu-vig/ezflow
cd ezflow/
python setup.py install

From PyPI

pip install ezflow

---

Models supported

• DICL
• DCVNet (1 checkpoint)
• FlowNetS
• FlowNetC (3 checkpoints)
• PWCNet (3 checkpoints)
• RAFT (3 checkpoints)
• VCN

Datasets supported

• AutoFlow
• FlyingChairs
• HD1K
• KITTI
• Kubric
• MPI Sintel
• SceneFlow Monkaa
• SceneFlow Driving
• SceneFlow FlyingThings3D
• SceneFlow FlyingThings3D subset

---

Results and Pre-trained checkpoints

• DCVNet | model config | paper

Training Dataset	Training Config	ckpts	Sintel Clean (training)	Sintel Final(training)	KITTI2015 AEPE	KITTI2015 F1-all
FlyingThings3DSubset + Monkaa + Driving	config	download	1.90	3.35	4.75	23.41%

• FlowNetC | model config | arXiv

Training Dataset	Training Config	ckpts	Sintel Clean (training)	Sintel Final(training)	KITTI2015 AEPE	KITTI2015 F1-all
Chairs	config	download	3.41	4.94	14.84	54.23%
Chairs -> Things	config	download	2.93	4.48	12.47	45.89%
Kubric	config	download	3.57	3.96	12.11	36.35%

• PWC-Net | model config | arXiv

Training Dataset	Training Config	ckpts	Sintel Clean (training)	Sintel Final(training)	KITTI2015 AEPE	KITTI2015 F1-all
Chairs	config	download	3.5	4.73	17.81	51.76%
Chairs -> Things	config	download	2.06	3.43	11.04	32.68%
Kubric	config	download	3.08	3.31	9.83	21.94%

• RAFT | model config | arXiv

Training Dataset	Training Config	ckpts	Sintel Clean (training)	Sintel Final(training)	KITTI2015 AEPE	KITTI2015 F1-all
Chairs	config	download	2.23	4.56	10.45	38.93%
Chairs -> Things	config	download	1.66	2.75	5.01	16.87%
Kubric	config	download	2.12	2.54	6.01	17.35%

---

Additional Information

• KITTI dataset has been evaluated with a center crop of size 1224 x 370.
• FlowNetC and PWC-Net uses padding of size 64 for evaluating the KITTI2015 dataset.
• RAFT and DCVNet uses padding of size 8 for evaluating the Sintel and KITTI2015 datasets.

---

References

• RAFT
• DICL-Flow
• PWC-Net
• FlowNetPytorch
• VCN
• detectron2
• CorrelationLayer
• ptflow

Pixels icon by Icons8