HOT3D Toolkit

This repository hosts an official toolkit for HOT3D, an egocentric dataset for 3D hand and object tracking.

The toolkit offers:

• An API for downloading and using the full HOT3D dataset provided in a VRS-based format (see tutorial).
• An API for HOT3D-Clips which is a curated HOT3D subset provided in the Webdataset format and used in BOP Challenge 2024 and Multiview Egocentric Hand Tracking Challenge.

Resources:

• HOT3D homepage
• Download full HOT3D dataset from projectaria.com
• Download HOT3D-Clips from Hugging Face
• Read HOT3D paper

The following instructions are relevant if you want to use the full HOT3D dataset in the VRS-based format. For HOT3D-Clips, please refer to the dedicated page.

Step 1: Install the downloader

This Python repository can be used with Pixi and Conda environments and can run on:

• x64 Linux distributions of:
  • Fedora 36, 37, 38
  • Ubuntu jammy (22.04 LTS) and focal (20.04 LTS)
• Mac Intel or Mac ARM-based (M1) with MacOS 11 (Big Sur) or newer

Python 3.9+ (3.10+ if you are on Apple Silicon).

Using Pixi

Pixi is a package management tool for developers. Developers can install libraries and applications in a reproducible way, which makes it easier to install and use a Python environment for the HOT3D API.

# 1. Install pixi
curl -fsSL https://pixi.sh/install.sh | bash

# 2. Checkout this repository
git clone https://github.com/facebookresearch/hot3d.git
cd hot3d/hot3d

# 3. Call `pixi install` to setup the environment
pixi install

# 4. (Optional) Install the third-party dependencies required for hands, by reviewing and accepting the licenses provided on the corresponding third-party repositories
pixi run setup_hands

A quick introduction to the PIXI environment

• Execute Pixi environment commands from within the hot3d folder.
• The Pixi environment is located in the .pixi folder.
• Activate the Pixi HOT3D environment by using the command pixi shell.
• Exit the Pixi HOT3D environment by typing exit.
• Remove the environment by executing the command rm -rf .pixi.

Using Conda

Conda is a package manager used for managing software environments and dependencies in data science and scientific computing.

# 1. Install conda -> https://conda.io/projects/conda/en/latest/user-guide/getting-started.html
# 2. Create your environment
conda create --name hot3d
conda activate hot3d

# 2. Install dependencies
python3 -m ensurepip
python3 -m pip install projectaria_tools==1.5.1 torch requests rerun-sdk==0.16.0
python3 -m pip install vrs
python3 -m pip install matplotlib

# 3. (Optional) Install the third-party dependencies required for hands by reviewing and accepting the licenses provided on the corresponding third-party repositories
python3 -m pip install 'git+https://github.com/vchoutas/smplx.git'
python3 -m pip install 'git+https://github.com/mattloper/chumpy'

A quick introduction to the CONDA environment

• Activate the Conda HOT3D environment by executing conda activate hot3d.
• Exit the Conda HOT3D environment using conda deactivate.
• Remove the Conda HOT3D environment by executing conda remove --name hot3d --all.

Step 2: Sign up and get the download links file

1. Review the HOT3D license agreement.
   • Examine the specific licenses applicable to the data types you wish to use, such as Sequence, Hand annotations, and 3D object models.
2. Go to the HOT3D website and sign up.
   • Scroll down to the bottom of the page.
   • Enter your email and select Access the Datasets.
3. The HOT3D page will be refreshed to contain instructions and download links
   • The download view is ephemeral, keep the tab open to access instructions and links
   • Download links that last for 14 days
   • Enter your email again on the HOT3D main page to get fresh links
4. Select the Download button for any of the data types:
   • “Download the HOT3D Aria Dataset"
   • "Download the HOT3D Quest Dataset"
   • "Download the HOT3D Assets Dataset"
5. These will swiftly download JSON files with urls that the downloader will use

Step 3: Download the data

Use the HOT3D downloader to download some, or all of the data.

# 1. Activate your environment (assuming from the hot3d folder):
# conda: conda activate hot3d
# pixi: pixi shell

# 2. Go to the hot3d/data_downloader directory
cd hot3d/data_downloader
mkdir -p ../dataset

# 3. Run the dataset downloader
# Download HOT3D Object Library data
python3 dataset_downloader_base_main.py -c Hot3DAssets_download_urls.json -o ../dataset --sequence_name all

# Download one HOT3D Aria data sequence
python3 dataset_downloader_base_main.py -c Hot3DAria_download_urls.json -o ../dataset --sequence_name P0003_c701bd11 --data_types all
# Type answer `y`

# Download one HOT3D Quest data sequence
python3 dataset_downloader_base_main.py -c Hot3DQuest_download_urls.json -o ../dataset --sequence_name P0002_1464cbdc --data_types all
# Type answer `y`

Tip: To download all sequences in a download links JSON file (such as the HOT3D Object Library data in step 3), pass sequence_name as 'all'.

Step 4: Run the dataset viewer

Viewing objects and headset pose trajectory

python3 viewer.py --sequence_folder <PATH>/hot3d_dataset/P0003_c701bd11 --object_library_folder <PATH>/hot3d_dataset/assets/

When using pixi, you can directly launch the viewer without explicitly activating the environment by using the following command:

pixi run viewer --sequence_folder <PATH>/hot3d_dataset/P0003_c701bd11 --object_library_folder <PATH>/hot3d_dataset/assets/

Using hand annotations

Hand pose annotations in HOT3D are provided in the UmeTrack and MANO formats. Both hand poses annotation are accessible in the API by using either the mano_hand_data_provider, umetrack_hand_data_provider property once the Hot3dDataProvider is initialized. In order to choose the representation for the viewer, use the following:

UmeTrack

python3 viewer.py --sequence_folder <PATH>/hot3d_dataset/P0003_c701bd11--object_library_folder <PATH>/hot3d_dataset/assets --hand_type UMETRACK

MANO

Hand annotations in the MANO format can be downloaded after accepting their license agreement.

• HOT3D only requires the MANO_RIGHT.pkl and MANO_LEFT.pkl files for loading and rendering of hand poses. These files can be obtained from the mano_v1_2.zip file located in the "Models & Code" section of the MANO website. After downloading, extract the zip file to your local disk, and the *.pkl files can be found at the following path: mano_v1_2/models/.

python3 viewer.py --sequence_folder <PATH>/hot3d_dataset/P0003_c701bd11 --object_library_folder <PATH>/hot3d_dataset/assets --mano_model_folder <PATH>/mano_v1_2/models/  --hand_type MANO

Step 5: Run the python notebook tutorial

# Assuming you have downloaded the Aria `P0003_c701bd11` sequence and the object library above.
#
# Install Jupyter Notebook for your environment:
python3 -m pip install Jupyter
# Run Jupyter and open the notebook (conda)
jupyter notebook ./HOT3D_Tutorial.ipynb
# Run Jupyter and open the notebook (pixi, use a direct path to ensure jupyter will take the right python path)
.pixi/envs/default/bin/jupyter notebook ./HOT3D_Tutorial.ipynb

Step6: Mastering the dataset and its API

Using the HOT3D dataset API

• Please refer to our notebook tutorial. The notebook explains how to instantiate a Hot3dDataProvider from a HOT3D sequence folder and how to use its API to retrieve each data modality (images, GT hand & object poses, ...).

Using "mask files"

HOT3D utilizes mask files to identify and flag specific timestamps based on a particular property. For instance, we can use masks to mark specific timestamps in a camera stream as having inaccurate object pose (as determined through manual QA) or as having an over-saturated image. These masks can be combined using logical AND/OR operations to create a custom mask that meets the end user's requirements.

Here is a list of the exported mask files:

File	Description
mask_object_pose_available.csv	True if all dynamic objects in the scene have a valid pose, and false if even one object is missing a pose. The accuracy of the poses is not evaluated.
mask_hand_pose_available.csv	True if both right and left hands have a valid pose, and false if even one hand is missing a pose. The accuracy of the poses is not evaluated.
mask_headset_pose_available.csv	True if the headset has a valid pose, and false otherwise. The accuracy of the poses is not evaluated.
mask_object_visibility.csv	True if at least one object is visible in the camera frame, false otherwise. Visibility is computed based on the pose, and accuracy of the pose is not evaluated.
mask_hand_visible.csv	True if at least one hand is visible in the camera frame, false otherwise. Visibility is computed based on the pose, and accuracy of the pose is not evaluated.
mask_good_exposure.csv	True if sufficient pixels on the hand and dynamic objects visible in the frame are not over-exposed, false otherwise. Visibility is computed based on the pose, and accuracy of the pose is not evaluated.
mask_qa_pass.csv	True if manual QA indicated no issues with the pose of the objects and hands visible in a camera stream, false otherwise.

The masks are saved in csv format, with each row indicating the validity of the mask for a given timestamp and camera stream_id:

timestamp[ns],stream_id,mask
43986008190448,214-1,True
43986008283010,1201-2,True
43986008283023,1201-1,True
43986041551363,214-1,True
43986041643100,1201-2,True
43986041643113,1201-1,True
43986074878771,214-1,False
43986074971171,1201-2,False
43986074971184,1201-1,False
43986108202591,214-1,True

Here is how to load the mask files and combine them with a logical operator:

# Mask API example:
# Let's check that we have at least a hand and an object that is visible in a given stream
#

import os
from projectaria_tools.core.stream_id import StreamId
from data_loaders.loader_masks import combine_mask_data, load_mask_data, MaskData

sequence_folder = "<PATH>/hot3d_dataset/P0003_c701bd11/"

# Select the desired masks
example_mask_list = [
    # Use the masks that depicting the visibility status
    "masks/mask_hand_visible.csv",
    "masks/mask_object_visible.csv",
]

# Load the referred masks
mask_data_list = []
for it in example_mask_list:
    if os.path.exists(os.path.join(sequence_folder, it)):
        ret = load_mask_data(os.path.join(sequence_folder, it))
        mask_data_list.append(ret)

# Combine the masks (you can choose logical "and"/"or")
output = combine_mask_data(mask_data_list, "and")

# Get the number of frames where we can see at least a hand and an object
num_timestamps_with_at_least_a_hand_and_object_visible = output.num_true(StreamId("214-1"))
print(f"Number of frames containing at least a hand and object visible: {num_timestamps_with_at_least_a_hand_and_object_visible}")

total_frame_count = output.length(StreamId("214-1"))
print(f"Total sequence frames: {total_frame_count}")

Test/Train split

The sequences corresponding to the following participant ids are defining the TEST set, and so does not have any GT information TEST_SET_PARTICIPANT_ID = ["P0004", "P0005", "P0006", "P0008", "P0016", "P0020"]

i.e:

• sequence P0003_c701bd11 belong to the TRAIN set
• sequence P0004_a59ab32e belong to the TEST set

Note that the information is also shared in the metadata.json file under the field gt_available_status or can be accessed directly via the Hot3dDataProvider -> get_sequence_metadata API call.

License

• HOT3D API (aka. this repository) is released by Meta under the Apache 2.0 license
• HOT3D dataset is released under the HOT3D license agreement
  • Using hands annotation requires installation of SMPLX/MANO third-party dependencies, please review and agree to their license listed on their website.

Contributing

Go to Contributing and the Code of Conduct.