hkchengrex / scribble-to-mask Goto Github PK

View Code? Open in Web Editor NEW

86.0 4.0 14.0 174 KB

[CVPR 2021] MiVOS - Scribble to Mask module

Home Page: https://hkchengrex.github.io/MiVOS/

License: MIT License

Python 100.00%

computer-vision deep-learning pytorch segmentation interactive-segmentation cvpr2021

scribble-to-mask's Introduction

MiVOS (CVPR 2021) - Scribble To Mask

Ho Kei Cheng, Yu-Wing Tai, Chi-Keung Tang

[arXiv] [Paper PDF] [Project Page]

A simplistic network that turns scribbles to mask. It supports multi-object segmentation using soft-aggregation. Don't expect SOTA results from this model!

Overall structure and capabilities

	MiVOS	Mask-Propagation	Scribble-to-Mask
DAVIS/YouTube semi-supervised evaluation	❌	✔️	❌
DAVIS interactive evaluation	✔️	❌	❌
User interaction GUI tool	✔️	❌	❌
Dense Correspondences	❌	✔️	❌
Train propagation module	❌	✔️	❌
Train S2M (interaction) module	❌	❌	✔️
Train fusion module	✔️	❌	❌
Generate more synthetic data	✔️	❌	❌

Requirements

The package versions shown here are the ones that I used. You might not need the exact versions.

PyTorch 1.6.0
torchvision 0.7.0
opencv-contrib 4.2.0
davis-interactive (https://github.com/albertomontesg/davis-interactive)
gitpython for training
gdown for downloading pretrained models

Refer to the official PyTorch guide for installing PyTorch/torchvision. The rest can be installed by:

pip install opencv-contrib-python gitpython gdown

Pretrained model

Download and put the model in ./saves/. Alternatively use the provided download_model.py.

[OneDrive Mirror]

Interactive GUI

python interactive.py --image <image>

Controls:

Mouse Left - Draw scribbles
Mouse middle key - Switch positive/negative
Key f - Commit changes, clear scribbles
Key r - Clear everything
Key d - Switch between overlay/mask view
Key s - Save masks into a temporary output folder (./output/)

Known issues

The model almost always needs to focus on at least one object. It is very difficult to erase all existing masks from an image using scribbles.

Training

Datasets

Download and extract LVIS training set.
Download and extract a set of static image segmentation datasets. These are already downloaded for you if you used the download_datasets.py in Mask-Propagation.

├── lvis
│   ├── lvis_v1_train.json
│   └── train2017
├── Scribble-to-Mask
└── static
    ├── BIG_small
    └── ...

Commands

Use the deeplabv3plus_resnet50 pretrained model provided here.

CUDA_VISIBLE_DEVICES=0,1 OMP_NUM_THREADS=4 python -m torch.distributed.launch --master_port 9842 --nproc_per_node=2 train.py --id s2m --load_deeplab <path_to_deeplab.pth>

Credit

Deeplab implementation and pretrained model: https://github.com/VainF/DeepLabV3Plus-Pytorch.

Citation

Please cite our paper if you find this repo useful!

@inproceedings{cheng2021mivos,
  title={Modular Interactive Video Object Segmentation: Interaction-to-Mask, Propagation and Difference-Aware Fusion},
  author={Cheng, Ho Kei and Tai, Yu-Wing and Tang, Chi-Keung},
  booktitle={CVPR},
  year={2021}
}

Contact: [email protected]

scribble-to-mask's People

Contributors

Stargazers

Watchers

Forkers

killsking tiamat-tech joskid nnmhuy cvlinks trendingtechnology stevesrh cv-ip jjho1314 amitrana001 csvt32745 rami-alloush zivzone pochsarah

scribble-to-mask's Issues

AttributeError: Caught AttributeError in DataLoader worker process 0

Hello! I followed the instructions of the training command, it has thrown an error about AttributeError.

I put the static folder outside this repository as you mentioned. It is confusing that I can use the same datasets for the pretraining propagation module, the train.py in Mask-Propagation works fine.

git.exc.InvalidGitRepositoryError when running train.py

Hello! I followed the instruction of the training command, but it has thrown an error about GitRepositoryError.

I used command : CUDA_VISIBLE_DEVICES=0,1 OMP_NUM_THREADS=4 python -m torch.distributed.launch --master_port 1842 --nproc_per_node=2 train.py --id s2m --load_deeplab ./deeplab_resnet50/best_deeplabv3plus_resnet50_voc_os16.pth, and I have 2 GPUs.
Could you give me some suggestions?

questions about range_transform.py

Hello ! Why im_mean was set to (124, 116, 104)?

Question about Local Control Strategy

A simple but practical segmentation tool! I've read your paper, and it says that local control strategy is used in S2M. However, I don't find the local control step in this code. Why don't you provide it in this tool? Will local control make significant difference to the performance?

saving error

Hello! Thanks for sharing your code. When I run python interactive.py and want to save the masks, appeared following error.

Could you give me some suggestions?

About evaluation of the model

Hi,

thank you for the nice work.

I have a concern about the evaluation of the model. Because there is no validation set to pick the best model. It may has a potential overfitting problem. (Or what should the validation set for interactive segmentation look like? If there is a unified standard, it will be more helpful for everyone to compare their methods.)

In interactive object segmentation setting, is this setting popular? I am new here for the interactive segmentation. Wish to solve my concern, thank you.

DeepLabv3 pre-trained models

Hello,

I wanted to mention that in order to train S2M from scratch, using the deeplabv3_resnet50 pre-trained model provided in this repo, returns the following error: KeyError: 'classifier.classifier.0.convs.0.0.weight.
Meaning that the weights from this layer are not present in deeplabv3_resnet50. But using the deeplabv3plus_resnet50 from the same repo executes without errors.

Best!