MotionLLM: Understanding Human Behaviors from Human Motions and Videos

TL;DR

MotionLLM introduces a unified framework that combines video and motion data for improved human behavior understanding, captioning, and reasoning, supported by a new dataset and benchmark for comprehensive evaluation.

Contribution

It presents a novel joint video-motion modeling approach, a large multi-modal dataset MoVid, and a benchmark MoVid-Bench for enhanced human behavior analysis.

Findings

MotionLLM outperforms existing models in captioning and reasoning tasks.

The combined use of video and motion data improves understanding accuracy.

Extensive experiments validate the effectiveness of the proposed framework.

Abstract

This study delves into the realm of multi-modality (i.e., video and motion modalities) human behavior understanding by leveraging the powerful capabilities of Large Language Models (LLMs). Diverging from recent LLMs designed for video-only or motion-only understanding, we argue that understanding human behavior necessitates joint modeling from both videos and motion sequences (e.g., SMPL sequences) to capture nuanced body part dynamics and semantics effectively. In light of this, we present MotionLLM, a straightforward yet effective framework for human motion understanding, captioning, and reasoning. Specifically, MotionLLM adopts a unified video-motion training strategy that leverages the complementary advantages of existing coarse video-text data and fine-grained motion-text data to glean rich spatial-temporal insights. Furthermore, we collect a substantial dataset, MoVid, comprising diverse videos, motions, captions, and instructions. Additionally, we propose the MoVid-Bench, with carefully manual annotations, for better evaluation of human behavior understanding on video and motion. Extensive experiments show the superiority of MotionLLM in the caption, spatial-temporal comprehension, and reasoning ability.