Finetuning Airbert on Downstream VLN Tasks

This repository stores the codebase for finetuning Airbert on downstream VLN tasks including R2R and REVERIE. The code is based on Recurrent-VLN-BERT. We acknowledge Yicong Hong for releasing the Recurrent-VLN-BERT code.

Prerequisites

1. Follow instructions in Recurrent-VLN-BERT to setup the environment and download data.

For REVERIE task, we use the same object features (REVERIE_obj_feats.pkl) as Recurrent-VLN-BERT for fair comparison. The pretrained Airbert can be found here.

2. Download the trained models.

REVERIE

Inference

To replicate the performance reported in our paper, load the trained models and run validation:

bash scripts/valid_reverie_agent.sh 0

Training

To train the model, simply run:

bash scripts/train_reverie_agent.sh 0

R2R

Inference

To replicate the performance reported in our paper, load the trained models and run validation:

bash scripts/valid_r2r_agent.sh 0

Training

To train the model, simply run:

bash scripts/train_r2r_agent.sh 0

Citation

Please cite our paper if you find this repository useful:

@misc{guhur2021airbert,
  title ={{Airbert: In-domain Pretraining for Vision-and-Language Navigation}},
  author={Pierre-Louis Guhur and Makarand Tapaswi and Shizhe Chen and Ivan Laptev and Cordelia Schmid},
  year={2021},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
}