input image0, estimated flow at each 5 scale, ground truth flow, input image1
- NVIDIA/flownet2-pytorch: framework, data transformers, loss functions, and many details about flow estimation.
- nameloss-Chatoyant/PWC-Net_pytorch: Referenced implmentation.
Working confirmed. I hope this helps you.
Unofficial implementation of CVPR2018 paper: Deqing Sun et al. "PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume". arXiv
-
Requirements
- Python 3.6+
- PyTorch 0.4.0 (mainly in in data handling)
- TensorFlow 1.8
-
model_3000epoch/model_3007.ckpt
is fully trained by SintelClean dataset.
# Training from scratch
python train.py --dataset SintelClean --dataset_dir path/to/MPI-Sintel-complete
# Start with learned checkpoint
python train.py --dataset SintelClean --dataset_dir path/to/MPI-Sintel-complete --resume model_3000epoch/model_3007.ckpt
After running above script, utilize GPU-id is asked, (-1:CPU). You can use other learning configs (like --n_epoch
or --batch_size
) see all arguments in train.py
, regards.