360-visualization

IA LAB PUC Research on VLN task, visualization fo 360° indoor environments on Matterport simulator.

We take the Matterport metadata and make it readily available to enhance tasks built on this dataset, such as the Visual and Language Navigation (VLN) task.

We also create functions for getting new crafted instructions (instructions based on rules with the best objects and viewpoints from the environment recommended for the agent to navigate) and the function for extracting these objects.

Fig. 1: Objects and reachable viewpoints visualization

Fig. 2: Crafted instruction generated with retrieved objects and viewpoints

Authors:

• Benjamin Earle (https://www.linkedin.com/in/benjamin-earle/)
• Joaquin Ossandón (https://www.linkedin.com/in/joaquin-ossandon/)

---

Repository distribution

├── LICENSE
├── README.md
├── docs/
├── img/
├── connectivity/ 👉🏽 Folder containing connectivity graphs of each scan/house
├── metadata_parser/
│   ├── house_cache/ 👉🏽 Folder containing serialized house metadata (with Pickle)
│   │   ├── 17DRP5sb8fy.pickle
│   │   ├── 1LXtFkjw3qL.pickle
│   │   ├── {...}.pickle
│   └── parse_house_segmentations.py 👉🏽 Script for generating house metadata pickles
└── slurm_scripts/ 👉🏽 Scripts for running repository scripts on Slurm (specifically on IA Lab Cluster)
    ├── jupyter_on_slurm.sh
    └── python_on_slurm.sh
├── data_explorer.ipynb 👉🏽 Explore the metadata of Matterport simulator, by house and viewpoint nodes.
├── simulator_visualization.ipynb 👉🏽 Jupyter Notebook to visualize objects coordinates and reachable viewpoints on 360° views, from Simulator
├── crafted_instructions_generator.ipynb 👉🏽 Jupyter Notebook to generate crafted instructions and visualize with panoramas
├── utils.py 👉🏽 Python file with common functions for visualizing and computing graph related stuff
├── objects_and_crafted_instructions_utils.py 👉🏽 Python file with functions for ranking recommended objects and generate crafted instructions
├── requirements.txt 👉🏽 All Python libraries for running the project

---

Installation

If you are member of the IA Lab (PUC), it is already installed on the servers, so the next steps are not necessary unless you would like to run on local.

We recommend following @peteanderson80 R2R paper guide for a successfully installation of Matterport simulator and all requirements for visualizing.

Some advices:

• Strongly recommended to follow the Docker setup.
• When running cmake, use cmake -DEGL_RENDERING=ON .., because we will use downloaded images. Also, this allows for off-screen rendering in the GPU.

As it says, you also will need to download the dataset. For object visualization you will need to follow this guide.

For parsing and using house metadata, see the metadata parser docs.

---

Run

You can run the notebooks with jupyter notebook command.

Please this guide for use of the object visualization.

---

(IA Lab only) Run on remote host

For local forwarding on remote server use

ssh -L localhost:8889:scylla:{PORT} <user>@<host>

where PORT is the port exposed by running the Jupyter Notebook.

If you are connecting to IA LAB Cluster, use:

sbatch /home/mrearle/repos/360-visualization/slurm_scripts/jupyter_on_slurm_gpu.sh
# or
sbatch /home/jiossandon/repos/360-visualization/slurm_scripts/jupyter_on_slurm_gpu.sh

you can see the exposed port by using cat {PATH}.

and cancel after not use

scancel <process ID SLURM> # you can review the ID with `sq`

Contributing

Please contact us if you want to contribute, use objects and crafted instructions to improve a vision-language model or anything else :)