Multi-agent Long-term 3D Human Pose Forecasting via
Interaction-aware Trajectory Conditioning

Jaewoo Jeong*, Daehee Park*, Kuk-Jin Yoon
Visual Intelligence Lab, KAIST
CVPR 2024 Highlight
*Indicates Equal Contribution
Paper Code Supplementary
MY ALT TEXT

Trajectory2Pose: a coarse-to-fine approach for human pose forecasting in complex scenes.
Our method first forecasts global trajectories (coarse), upon which local poses (fine) are conditionally predicted.

Methodology

MY ALT TEXT
  • - Coarse global trajectories are forecasted from past motion (trajectory + pose)
  • - Fine local motion forecasted with conditioning on predicted future trajectory
  • - Complementary forecasting of Global and Local motion yields comprehensive proficiency.

Comparison of results

Comparison of prediction results on CMU-Mocap (UMPM) dataset. We compare results with MRT, JRT, and TBIFormer. BLACK: GT (1s+2s), COLOR: Prediction (2s)

T2P (Ours)

MRT / 2021 NeurIPS

JRT / 2023 ICCV

TBIFormer / 2023 CVPR

More results

We show more results of our T2P model on CMU-Mocap (UMPM) samples. BLACK: GT (1s+2s), COLOR: Prediction (2s)

JRDB-GMP dataset

We also parse a new dataset to train forecasting model on complex (long-term: 2s+, multi-agent: 3+), real world data. Parsed from JRDB dataset, which is acquired from a mobile robot on campus. (RGB+LiDAR)

BibTeX


        @inproceedings{jeong2024multi,
          title={Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning},
          author={Jeong, Jaewoo and Park, Daehee and Yoon, Kuk-Jin},
          booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
          pages={1617--1628},
          year={2024}
        }