Skip to content

[NeurIPS 2020] Displacement-Invariant Matching Cost Learning for Accurate Optical Flow Estimation

License

Notifications You must be signed in to change notification settings

jytime/DICL-Flow

Repository files navigation

DICL-Flow

This repository contains the code for our NeurIPS 2020 paper Displacement-Invariant Matching Cost Learning for Accurate Optical Flow Estimation.

@article{wang2020displacement,
  title={Displacement-Invariant Matching Cost Learning for Accurate Optical Flow Estimation},
  author={Wang, Jianyuan and Zhong, Yiran and Dai, Yuchao and Zhang, Kaihao and Ji, Pan and Li, Hongdong},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}

Requirements

PyTorch 1.1, CUDA 9.0, and other dependencies (e.g., torchvision, tqdm). You can create a virtual environment and use pip to install the requirements. You may also install the dependencies manually.

conda create -n dicl python=3.6
conda activate dicl
pip install -r requirements.txt

Dataset

We conduct experiments on the FlyingChairs, FlyingThings3D, MPI Sintel, and KITTI optical flow datasets.

My dataset structure is listed below. You can put data at any place and set the --data flag of main.py correspondingly.

├── dataset
    ├── FlyingChairs_release
        ├── data
    ├── FlyingThings3D_subset
        ├── train
        ├── val
    ├── Sintel
        ├── training
        ├── test
    ├── KITTI2012
        ├── training
        ├── testing
    ├── KITTI2015
        ├── training
        ├── testing

Pretrained Weights

Pretrained models can be downloaded from google drive.

Training

As discussed in our paper, the training on the FlyingChairs dataset is divided into three phases for a good initialization (only single phase on other three datasets). We assume using eight GPUs while empirically the number of GPUs does not affect the final performance. You can use the commands like below:

# Chair 0 phase, only simple augmentation is applied in the first phase.
# Please note each epoch contains around 350 iteration on FlyingChairs, 22232/64
python main.py -b 64 --lr 0.001 --epochs 120 --exp_dir dicl0_chair --cfg cfgs/dicl0_chair.yml \
--data /Path/To/FlyingChairs --dataset flying_chairs

# Chair 1 phase, using context network now.
python main.py -b 64 --lr 0.001 --epochs 120 --exp_dir dicl1_chair --cfg ../cfgs/dicl1_chair.yml \
--pretrained /Path/To/dicl0/checkpoint_best.pth.tar --data /Path/To/FlyingChairs --dataset flying_chairs

# Chair 2 phase, using Displacement Aware Projection layer now.
# Drop the learning rate by half at epoch 10.
python main.py -b 64 --lr 0.001 --epochs 120 --exp_dir dicl2_chair --cfg ../cfgs/dicl2_chair.yml --milestones 10 \
--pretrained /Path/To/dicl1/checkpoint_best.pth.tar --data /Path/To/FlyingChairs --dataset flying_chairs

For other datasets, just use one command, e.g.,

python main.py -b 16 --lr 0.00025 --epochs 50 --pretrained /Path/To/dicl2/checkpoint_best.pth.tar
--exp_dir dicl3_thing --cfg cfgs/dicl3_thing.yml --data /Path/To/FlyingThings --dataset flying_things

Evaluation

You can evaluate a trained model like below, where -e indicates evaluation:

# Sintel Dataset
python main.py -b 1 -e --pretrained pretrained/ckpt_sintel.pth.tar --cfg cfgs/dicl4_sintel.yml \
--data /Path/To/Sintel/Dataset --exp_dir /Path/To/Save/Log --dataset mpi_sintel_clean 

# KITTI Dataset
python main.py -b 1 -e --pretrained pretrained/ckpt_kitti.pth.tar --cfg cfgs/dicl5_kitti.yml \
--data /Path/To/KITTI/Dataset --exp_dir /Path/To/Save/Log --dataset KITTI

Acknowledgment

Our codes were developed on the basis of FlowNetPytorch and VCN. After the paper acceptation, we updated the part of Dataloader, learning from RAFT. Thanks a lot for their excellent works.

About

[NeurIPS 2020] Displacement-Invariant Matching Cost Learning for Accurate Optical Flow Estimation

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages