Loco3D: Indoor Multiuser Locomotion 3D Dataset

Kojiro Takeyama, Yimeng Liu, Misha Sra

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Human trajectory synthesis, Indoor, Dataset, Multi-user, 3D, Virtual reality

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: In the context of human-AI interaction, modeling human actions is a critical and challenging endeavor, with locomotion being a particularly fundamental behavior for AI agents to understand. Modeling human trajectories in complex indoor scenes, such as the home environment, requires an understanding of how humans interact with their surroundings and other humans. These interactions are influenced by a range of factors, including the geometry and semantics of the scene, the socio-cultural context, and the task each human needs to perform. Previous research has shared datasets containing human motion and scene structure in indoor scenes, but these datasets are limited in scale due to the difficulty and time required to collect data at different locations. To solve the scale problem, we propose to use a virtual reality (VR) system to build a human motion dataset. Specifically, we present Loco3D, a dataset of multi-person interactions in over 100 different indoor VR scenes, including 3D body pose data and highly accurate spatial information. The dataset can be used for building AI agents that operate in indoor environments, such as home robots, or to create virtual avatars for games or animations that mimic human movement and posture. With an initial evaluation, we demonstrate that models trained with our dataset have improved multi-person trajectory synthesis performance on real-world data.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6448