Target-Driven Structured Transformer Planner for Vision-Language Navigation
This is the official implementation of MM'22 (ACM International Conference on Multimedia) paper (accepted as oral): Target-Driven Structured Transformer Planner for Vision-Language Navigation.
Citation
@inproceedings{zhao2022target,
title={Target-Driven Structured Transformer Planner for Vision-Language Navigation},
author={Zhao, Yusheng and Chen, Jinyu and Gao, Chen and Wang, Wenguan and Yang, Lirong and Ren, Haibing and Xia, Huaxia and Liu, Si},
booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
pages={4194--4203},
year={2022}
}
Installation
Please refer to this repo for installation guides.
Training & Inference
To train on R2R:
cd finetune_src
python ./scripts/run_r2r.sh
To train on REVERIE:
cd finetune_src
python ./scripts/run_reverie.sh
To test on R2R:
cd finetune_src
python ./scripts/test_r2r.sh
To test on REVERIE:
cd finetune_src
python ./scripts/test_reverie.sh
Note that
- Some file paths in the code may require slight adaptation according to your local environment.
- To achieve the RGS and RGSPL of REVERIE in the paper, you need to train a separated ViL-BERT on the REVERIE training split and perform the referring part at the end of the navigation. Since the referring part is not the contribution of our paper and is easy to tune, we do not plan to release that part of code.
Acknowledgement
This code is based on HAMT. We appreciate their great contribution to the community.