Official implementation of 'Learning Motion-Appearance Co-Attention for Zero-Shot Video Object Segmentation', ICCV-2021

in Pytorch

Paper

• Python 3.6/3.7
• Pytorch
• Torchvision
• Numpy
• Pillow
• Opencv-python (cv2)
• Scikit-image
• Joblib
• Tqdm
• Cython
• Pydensecrf

If you have some problems installing pydensecrf, you can visit pydensecrf for more information.

In datasets/dataloader_list:

adaptor_dataset.py and adaptor_dataset_.py are used to assemble the DAVIS and Youtube-VOS datasets to build the model dataset, in practice we only use DAVIS for training.

davis.py and tyb_vos.py represent the data set construction rules, which can be modified to get the desired data list and stored in the cache (e.g. datasets/DAVIS/cache/tran2016.pkl and datasets/DAVIS/cache/val2016.pkl).

custom_transforms_f.py and transform.py contain some functions for data augmentation.

Avaliable flow maps from PWCNet can be find in baidu(mvuv)

• Set the training sets in utils/args.py and train.py (Line 163 - Line 170).
• Run train.py

  python train.py

• Test

  Please use test.py to generate results.

  python test.py

  Then you can get the predicted results in the output/davis16/AMCNet/results folder.

• CRF

  At the same time, CRF can be further performed on the predicted results, and the relevant code is in utils/apply_densecrf_davis.py.

  python utils/apply_densecrf_davis.py

  Then you can get the predicted results in the output/davis16/AMCNet/results_crf folder.

  Line 18: dataset images folder

  Line 19: prediction mask folder

  Line 20: crf prediction folder

Please use the files in the EVALVOS folder to measure metrics.

Taking test_for_davis.py as an example:

Line 13: Setup db_info.yml

Line 14: Set the folder of groundtruth

Line 254: Set the folder of images

Line 255 & 256: Whether to discard the first frame and the last frame

Line 257: Save output in .pkl format

python test_for_davis.py

I have provided some tools, which can be seen in README