Home

Awesome

Learning the What and How of Annotation in Video Object Segmentation

Created by Thanos Delatolas, Vikcy Kalogeiton, Dim P. Papadopoulos

EVA-VOS

[Paper (WACV 2024)] [Project page] [Extended Abstract (ICCV-W 2023)]

Installation

conda env create -f environment.yaml

Data

Download the data with python download_data.py. The data should be arranged with the following layout:

data
├── DAVIS_17
│   ├── Annotations
│   ├── ImageSets
│   └── JPEGImages
│           
├── MOSE
│   ├── Annotations
│   ├── ImageSets
│   └── JPEGImages

The script download_data.py also creates the train/val/test splits in MOSE, as discussed in the paper. If qdown denies access to the MOSE dataset, you can manually download MOSE from here and place it in the directory: ./data/MOSE/

Download weights

Download the model-weights with python download_weights.py. The weights should be arranged with the following layout:

model_weights
├── mivos
│   └── stcn.pth
│   └── fusion.pth
├── qnet
│   └── qnet.pth
├── rl_agent
│   └── model.pth
├── sam
│   └── sam.pth

We provide the weights of MiVOS trained only on YouTube-VOS. If you wish to replicate the training process, please refer to the original repository.

Training

Experiments

The script eval_annotation_method.py is used to execute all annotation methods. The script scripts/eval.sh can be used to run all the experiments. Finally, the scripts vis/frame_selection.py and vis/full_pipeline.py are used to plot the results obtained from the experiments conducted. To speed up the process, it is recommended to run the experiments simultaneously on multiple GPUs.

Citation

@inproceedings{delatolas2024learning,
  title={Learning the What and How of Annotation in Video Object Segmentation},
  author={Thanos Delatolas and Vicky Kalogeiton and Dim P. Papadopoulos},
  year={2024},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)}
}