Next: Setup simulator

To download data, run

$ bash download.sh

Important: The ImageNet features are not maintained by us, so if the link would not work in the future, please visit: https://github.com/peteanderson80/Matterport3DSimulator#precomputing-resnet-image-features and click on "ResNet-152-imagenet features [380K/2.9GB]" to get the latest link, or email an author of the VLN paper.

The following directories and files will be created:

• asknav and noroom: VNLA datasets. Please read our paper for more detail about each dataset.
• connectivity: environment graphs.
• img_features: precomputed image embeddings computed by ResNet pretrained on ImageNet.
• region_label.txt: room names.
• accepted_objects.txt: object labels.

Request access to the dataset here. The dataset is for non-commercial academic purposes only. Please read and agree to the dataset's terms and conditions and put this link in your project repo as requested by the dataset's creators.

Training and testing our models only require downloading the house_segmentations portion of the dataset. Unzip the files so that <some_folder>/v1/scans/<scanId>/house_segmentations/panorama_to_region.txt are present.

Running in graphics mode is still useful for debugging and visualizing the agent behavior. You need to download the matterport_skybox_images portion and unzip the files so that <some_folder>/v1/scans/<scanId>/matterport_skybox_images/*.jpg are present.

We provide the script unzip_matterport.sh in this directory to help you unzip the data:

$ bash unzip_matterport.sh $matter_root

where $matter_root is the Matterport3D dataset top folder where $matter_root/v1/scans/ is located.

Next: Setup simulator