This code implementation of TinyML 2021 symposium paper : ightweight Video Object Segmentation with Adaptive Template Attention Module and Temporal Consistency Loss .
We improve accuracy more by using better attention clue. The paper about new method will be uploaded soon.
python (>= 3.6)
numpy
pytorch (>= 1.8 )
torchvision
pillow
tqdm
imgaug
opencv
DAVIS : download here
/path/DAVIS
|-- Annotations/
|-- ImageSets/
|-- JPEGImages/
YouTubeVOS : download here
/path/ytvos2018
|-- train/
|-- train_all_frames/
|-- valid/
`-- valid_all_frames/
Saliency dataset : download here
/path/Saliency
|-- ECSSD/
|-- HKU-IS/
|-- MSRA10K/
Backbone | FLOP (G) | Param (M) | J & F 16 | J & F 17 | FPS | link |
---|---|---|---|---|---|---|
HRNet | 10.61 | 1.61 | 81.1 | 62.1 | 78.3 | Google Drive |
RN18 | 55.23 | 12.5 | 82.2 | 66.3 | 54.5 | Google Drive |
RN50 | 83.86 | 14.8 | 83.1 | 69.5 | 37.7 | Google Drive |
MobileNetV3 | 7.56 | 3.66 | 79.8 | 62.9 | 74.6 | Google Drive |
- Downlaod pre-trained HRNet in here and re-name to 'hrnet_w18_small_model_v1.pth'
- Put the weights at the path ['./nnWeight']
- Run train.py
- put the best file path [save_dir] and name [pth] in config setting
This codebase borrows the code and structure from official A-GAME repository. We are grateful to Facebook Inc. with valuable discussions.
The codebase is built based on following works
@article{park2020ttvos,
title={TTVOS: Lightweight Video Object Segmentation with Adaptive Template Attention Module and Temporal Consistency Loss},
author={Park, Hyojin and Venkatesh, Ganesh and Kwak, Nojun},
journal={arXiv preprint arXiv:2011.04445},
year={2020}
}
}