[ECCV 2022] Flow-Guided Transformer for Video Inpainting

[Paper] / [Codes] / [Demo] / [Project page]

This repository hosts the pretrained models of the following paper:

Flow-Guided Transformer for Video Inpainting<br> Kaidong Zhang, Jingjing Fu and Dong Liu<br> European Conference on Computer Vision (ECCV), 2022<br>

Details

There are three models in this repository, here are the details.

• lafc.pth.tar: The pretrained model of "Local Aggregation Flow Completion Network", which accepts a sequence of corrupted optical flows, and outputs the completed flows.
• lafc_single.pth.tar: The pretrained model of the single flow completion version of "Local Aggregation Flow Completion Network", it accepts one corrupted flow, and outputs one completed flow. (Only for the training of the FGT model)
• fgt.pth.tar: The pretrained model of "Flow Guided Transformer", which receives a sequence of corrupted frames and completed optical flows, and outputs the completed frames.

Besides the pretrained weights, we also provide the configuration files of these pretrained models.

• LAFC_config.yaml: The configuration file of lafc.pth.tar
• LAFC_single_config.yaml: The configuration file of lafc_single.pth.tar
• FGT_config.yaml: The configuration file of fgt.pth.tar

Deployment

Download this repository to the base directory of the codes (please download that at the github page), and run "bash deploy.sh" to form the models and the cofiguration files.

After the step above, you can skip the step 1~3 in the quick start section in the github page and run the object removal demo directly.