SIRNet

This is the PyTorch implementation for "Boosting Video Object Segmentation based on Scale Inconsistency", ICME, 2022. We present a refinement framework to boost the performance of pre-trained semi-supervised video object segmentation models. Project page

Installation:

Download the DAVIS 16 and DAVIS 17 dataset
Edit the paths in mypath.py

Offline training:

Training network on DAVIS 16 to get a good initialization
```
python train_p.py
```
Edit the path in train_p_ms.py with the path to pretrained model and start offline training with our multi-scale context aggregation module with scale inconsistency-based variance regularization.
```
python train_p_ms.py
```

Online learning:

For online learning on DAVIS16 datasets, please edit the path in train_o_ms_so.py with the path to your offlined trained model and start online learning
```
python train_o_ms_so.py
```
For online learning on DAVIS17 datasets, please edit the path in train_o_ms_mo.py with the path to your offlined trained model and start online learning
```
python train_o_ms_mo.py
```

Online Adaptation:

For online adaptation on DAVIS16 datasets, please edit the path in train_ada_so.py with the path to your online learned model and start online adaptation
```
python train_ada_so.py
```
For online adaptation on DAVIS17 datasets, please edit the path in train_ada_mo.py with the path to your online learned model and start online adaptation
```
python train_ada_so.py
```

Visualization of Inter-frame adaptation loss

In ./Results, we put the visualization of our inter-frame adaptation loss. You can also generate this loss by running