MaskFlownet-Pytorch

Unofficial PyTorch implementation of MaskFlownet (https://github.com/microsoft/MaskFlownet).

Tested with:

  • PyTorch 1.5.0
  • CUDA 10.1

Install

The correlation package must be installed first:

cd model/correlation_package
python setup.py install

Inference

Right now, I implemented the inference script for KITTI 2012/2015, MPI Sintel and FlyingChairs.

python predict.py CONFIG -c CHECKPOINT --dataset_cfg DATASET -f ROOT_FOLDER [-b BATCH_SIZE]

For example:

  • python predict.py MaskFlownet.yaml -c 5adNov03-0005_1000000.pth --dataset_cfg sintel.yaml -f ./SINTEL -b 4
  • python predict.py MaskFlownet.yaml -c 8caNov12-1532_300000.pth --dataset_cfg kitti.yaml -f ./KITTI -b 4
  • python predict.py MaskFlownet_S.yaml -c 771Sep25-0735_500000.pth --dataset_cfg chairs.yaml -f ./FLYINGCHAIRS -b 4
  • python predict.py MaskFlownet_S.yaml -c dbbSep30-1206_1000000.pth --dataset_cfg sintel.yaml -f ./SINTEL -b 4

Differences with the original implementation

The results are slightly different from the original implementation:

Checkpoint Network Implementation KITTI2012 KITTI2015 Sintel Clean Sintel Final FlyingChairs
771Sep25 MaskFlownet_S

Original AEPE:
PyTorch AEPE:

4.12
4.18

11.52
11.82

3.38
3.38

4.71
4.70

1.84
1.83
dbbSep30 MaskFlownet_S

Original AEPE:
PyTorch AEPE:

1.27
1.28

1.92
1.93

2.76
2.78

3.29
3.32

2.36
2.36
5adNov03 MaskFlownet

Original AEPE:
PyTorch AEPE:

1.16
1.18

1.66
1.68

2.58
2.59

3.14
3.17

2.23
2.23
8caNov12 MaskFlownet

Original AEPE:
PyTorch AEPE:

0.82
0.82

1.38
1.38

4.34
4.40

5.27
5.33

4.01
3.99

Examples

KITTI Original implementation:

original-implementation

KITTI This implementation:

this-implementation

Sintel Original implementation:

original-sintel

Sintel This implementation:

this-sintel

FlyingChairs Original implementation:

original-chairs

FlyingChairs This implementation:

this-chairs

Notes

If you use my implementation for training, it might happen that you encounter this error:

CUDA error: an illegal memory access was encountered

This is due to a bug in the torchvision implementation of deformable convolutions. (still present in version 0.7.0)

To solve it, you need to use the nightly version of torchvision.

GitHub

https://github.com/cattaneod/MaskFlownet-Pytorch