Skip to content

Code for ORAR Agent for Vision and Language Navigation on Touchdown and map2seq

License

Notifications You must be signed in to change notification settings

raphael-sch/map2seq_vln

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Code for ACL 2022 paper: Schumann and Riezler, "Analyzing Generalization of Vision and Language Navigation to Unseen Outdoor Areas "

Results for Model Weights in this Repository

Model TC SPD SED TC SPD SED
dev dev dev test test test
touchdown seen:
- no image 15.38 25.66 15.03 13.13 27.52 12.58
- 4th-to-last 30.05 11.12 29.46 29.60 11.79 28.89
- 4th-to-last no head. & no junc. 24.08 13.63 23.48 24.49 14.19 23.98
touchdown unseen:
- no image 11.50 25.89 10.72 9.62 27.71 9.04
- 4th-to-last 16.88 19.59 16.21 15.06 20.52 14.32
map2seq seen:
- no image 42.25 8.02 41.52 39.25 7.90 38.42
- pre-final 49.88 5.87 48.96 47.75 6.53 46.76
- pre-final no head. & no junc. 49.62 7.36 48.80 44.88 8.79 44.21
map2seq unseen:
- no image 27.88 10.71 27.05 30.25 11.52 29.02
- 4th-to-last 27.62 11.81 26.96 29.62 13.16 28.91
merged seen:
- no image 27.80 18.48 27.15 24.63 19.51 23.91
- 4th-to-last 38.20 9.18 37.42 36.22 9.55 35.42
merged unseen:
- no image 22.19 17.85 21.46 19.85 21.20 19.10
- 4th-to-last 25.31 15.17 24.24 24.10 16.48 23.46

Workflow without Images

(no need to download and preprocess panoramas)

Preparation

pip install -r requirements.txt

Inference and Evaluation

python vln/main.py --test True --dataset map2seq_unseen --config outputs/map2seq_unseen/noimage/config/noimage.yaml --exp_name noimage --resume SPD_best

Train from Scratch:

python vln/main.py --dataset touchdown_unseen --config configs/noimage.yaml --exp_name no_image

Workflow with Images

Panorama Preprocessing

Unfortunately we are not allowed to share the panorama images or the ResNet features derived from them. You have to request to download the images here: https://sites.google.com/view/streetlearn/dataset
Then change into the panorama_preprocessing/last_layer or panorama_preprocessing/fourth_layer folder and use the extract_features.py script.

Preparation

pip install -r requirements.txt

Test

python vln/main.py --test True --dataset touchdown_seen --img_feat_dir 'path_to_features_dir' --config link_to_config --exp_name 4th-to-last --resume SPD_best

The path_to_features_dir should contain the resnet_fourth_layer.pickle and resnet_last_layer.pickle file created in the pano preprocessing step.

Train from Scratch:

python vln/main.py --dataset touchdown_seen --img_feat_dir 'path_to_features_dir' --config configs/4th-to-last.yaml --exp_name 4th-to-last

References

Code based on https://github.com/VegB/VLN-Transformer
Touchdown splits based on: https://github.com/lil-lab/touchdown
map2seq splits based on: https://map2seq.schumann.pub
Panorama images can be downloaded here: https://sites.google.com/view/streetlearn/dataset

Citation

Please cite the following paper if you use this code:

@inproceedings{schumann-riezler-2022-analyzing,
    title = "Analyzing Generalization of Vision and Language Navigation to Unseen Outdoor Areas",
    author = "Schumann, Raphael and Riezler, Stefan",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    year = "2022",
    address = "Dublin, Ireland",
    pages = "7519--7532"
}

Releases

No releases published

Packages

No packages published

Languages