By Shengyu Zhao, Yilun Sheng, Yue Dong, Eric I-Chao Chang, Yan Xu.
@inproceedings{zhao2020maskflownet,
author = {Zhao, Shengyu and Sheng, Yilun and Dong, Yue and Chang, Eric I-Chao and Xu, Yan},
title = {MaskFlownet: Asymmetric Feature Matching with Learnable Occlusion Mask},
booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2020}
}
Feature warping is a core technique in optical flow estimation; however, the ambiguity caused by occluded areas during warping is a major problem that remains unsolved. We propose an asymmetric occlusion-aware feature matching module, which can learn a rough occlusion mask that filters useless (occluded) areas immediately after feature warping without any explicit supervision. The proposed module can be easily integrated into end-to-end network architectures and enjoys performance gains while introducing negligible computational cost. The learned occlusion mask can be further fed into a subsequent network cascade with dual feature pyramids with which we achieve state-of-the-art performance. For more details, please refer to our paper.
This repository includes:
Code has been tested with Python 3.6 and MXNet 1.5.
We follow the common training schedule for optical flow using the following datasets:
Please modify the paths specified in main.py
(for FlyingChairs), reader/things3d.py
(for FlyingThings3D), reader/sintel.py
(for Sintel), reader/kitti.py
(for KITTI 2012 & KITTI 2015), and reader/hd1k.py
(for HD1K) according to where you store the corresponding datasets. Please be aware that the FlyingThings3D dataset (subset) is still very large, so you might want to load only a relatively small proportion of it (see main.py
).
The following script is for training:
python main.py CONFIG [-dataset_cfg DATASET_CONFIG] [-g GPU_DEVICES] [-c CHECKPOINT, --clear_steps] [--debug]
where CONFIG
specifies the network and training configuration; DATASET_CONFIG
specifies the dataset configuration (default to chairs.yaml
); GPU_DEVICES
specifies the GPU IDs to use (default to cpu only), split by commas with multi-GPU support. Please make sure that the number of GPUs evenly divides the BATCH_SIZE
, which depends on DATASET_CONFIG
(BATCH_SIZE
are 8
or 4
in the given configurations, so 4
, 2
, or 1
GPU(s) will be fine); CHECKPOINT
specifies the previous checkpoint to start with; use --clear_steps
to clear the step history and start from step 0; use --debug
to enter the DEBUG mode, where only a small fragment of the data is read. To test whether your environment has been set up properly, run: python main.py MaskFlownet.yaml -g 0 --debug
.
Here, we present the procedure to train a complete MaskFlownet model for validation on the Sintel dataset. About 20% sequences (ambush_2, ambush_6, bamboo_2, cave_4, market_6, temple_2) are split as Sintel val, while the remaining are left as Sintel train (see Sintel_train_val_maskflownet.txt
). CHECKPOINT
in each command line should correspond to the name of the checkpoint generated in the previous step.
Pretrained models for step 2, 3, and 6 in the above procedure are given (see ./weights/
).
The following script is for inferring:
python main.py CONFIG [-g GPU_DEVICES] [-c CHECKPOINT] [--valid or --predict] [--resize INFERENCE_RESIZE]
where CONFIG
specifies the network configuration (MaskFlownet_S.yaml
or MaskFlownet.yaml
); GPU_DEVICES
specifies the GPU IDs to use, split by commas with multi-GPU support; CHECKPOINT
specifies the checkpoint to do inference on; use --valid
to do validation; use --predict
to do prediction; INFERENCE_RESIZE
specifies the resize used to do inference.
For example,
to do validation for MaskFlownet-S on checkpoint fffMar16
, run python main.py MaskFlownet_S.yaml -g 0 -c fffMar16 --valid
(the output will be under ./logs/val/
).
to do prediction for MaskFlownet on checkpoint 000Mar17
, run python main.py MaskFlownet.yaml -g 0 -c 000Mar17 --predict
(the output will be under ./flows/
).
For those who do not wish to train the model and would purely like to obtain flow images from a pretrained model on their own data, please use predict_new_data.py. You do not need to download any of the optical flow datasets to use predict_new_data.py, although you will have to additionally pip install flow_vis and moviepy. The functions provide a means to load a model and perform inference on a given pair of images or to obtain a series of flow images corresponding to the movement between component images of a given video without the need to download optical flow datasets. These can be called from another script or you can call the program from a terminal/Anaconda prompt like so:
to obtain a video composed of the flow images corresponding to input_video.mp4
, run python predict_new_data.py C:/Users/my_username/flow_video_filepath.mp4 MaskFlownet.yaml --video_filepath C:/Users/my_username/input_video.mp4 -g 0 -c 8caNov12
to obtain a flow image from 2 input images image_1.png
and image_2.png
, run python predict_new_data.py C:/Users/my_username/flow_image_filepath.png MaskFlownet.yaml --image_1 C:/Users/my_username/image_1.png --image_2 C:/Users/my_username/image_2.png -g 0 -c 8caNov12
We thank Tingfung Lau for the initial implementation of the FlyingChairs pipeline.