Transformer-UVOS
This is a repo offers a strong baseline for Unsupervised Video Object Segmentation(UVOS) using Transformer
Accuracy
With Res50 as backbone and image size (640x352) during training, using only single-scale for testing On DAVIS 2016(val):
| Method | J_mean | J_recall | J_decay | F_mean | F_recall | F_decay |
|---|---|---|---|---|---|---|
| Ours(Res50-FPN, 640x352) | 0.777 | 0.915 | 0.066 | 0.766 | 0.859 | 0.043 |
| Anchor-Diffusion(Res101-Deeplabv3, 854x480) | 0.782 | --- | --- | 0.771 | --- | --- |
Architecture
ResNet50-FPN + Transformer + Simple Decoder
Training
You need to first download the DAVIS dataset
The network is initialized using ResNet50-FPN pre-trained on COCO dataset, which can be downloaded from here.
cd model/transformer/davis.transformer.fpn.R50.random_sample/
sh run.sh
Acknowledgement
Idea inspired by Anchor-Diffusion.
Codes based on TorchSeg, Detectron2 and Detr
Contact
Feel free to contact me if you have any questions : yuming.du@enpc.fr