TSDTVOS

This is the implementation of TSDTVOS.
This code is based on MiVOS: [link].

Display

Dependencies

• PyTorch 1.7.1
• torchvision 0.8.2
• OpenCV 4.2.0
• progressbar

DataSets

• DAVIS
  download DAVIS2016 and DAVIS2017(480p) from [link]

path/to/DAVIS
├── 2016
│   ├── Annotations
│   └── ...
└── 2017
    ├── trainval
    │   ├── Annotations
    │   └── ...
    ├── test-dev
        ├── Annotations
        └── ...

• YouTubeVOS
  download YouTubeVOS2018 valid from [link]
  download YouTubeVOS2018 valid_all_frames from [link]
  download YouTubeVOS2019 valid from [link]
  download YouTubeVOS2019 valid_all_frames from [link]

path/to/YouTubeVOS
├── 2018
│   ├── all_frames
│   │   └── valid_all_frames
│   └── valid
└── 2019
    ├── all_frames
    │   └── valid_all_frames
    └── valid

Trained model

• Download pre-trained weights into ./saves
  [weights]

Code

• DAVIS-2016 validation set

python eval_davis_2016.py --davis_path [path/to/DAVIS-2016] --output [path/to/output]

• DAVIS-2017 validation set

python eval_davis.py --davis_path [path/to/DAVIS-2017] --split val --output [path/to/output]

• DAVIS-2017 test-dev set

python eval_davis.py --davis_path [path/to/DAVIS-2017] --split testdev --output [path/to/output] --conf_thr 0.4

• YouTubeVOS-2018 validation set

python eval_youtube.py --yv_path [path/to/YouTubeVOS-2018] --output [path/to/output]

• YouTubeVOS-2019 validation set

python eval_youtube.py --yv_path [path/to/YouTubeVOS-2019] --output [path/to/output]

Pre-computed Results

We provide pre-computed results for benchmark sets.

DAVIS-16-val
DAVIS-17-val
YouTube-VOS-18-valid
YouTube-VOS-19-valid
DAVIS-17-testdev