Research map / 3D vision / Human pose estimation

SimpliHuMoN: Simplifying Human Motion Prediction

NeurIPS 2026 · 3D vision › Human pose estimation · approach: Human motion prediction · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SimpliHuMoN, a unified decoder-only transformer with self-attention over past context and learnable future queries that handles pose-only, trajectory-only, and joint pose+trajectory prediction with winner-takes-all multi-hypothesis output, reporting SOTA or competitive results on Human3.6M, AMASS, ETH-UCY, SDD, MOCAP-UMPM, and 3DPW.

Problem

Human motion prediction has been split into specialized pose and trajectory forecasting models; holistic models often underperform on individual task benchmarks, so a single model that is competitive across pose, trajectory, and joint prediction is needed.

Approach

A single, minimal transformer decoder-only model concatenates encoded past context (root trajectory and/or relative pose tokens) with learnable future query tokens, applies self-attention over the whole sequence, and decodes K hypotheses with a winner-takes-all loss, with no task-specific modules.

Key results

• On MOCAP-UMPM, the wide model reaches APE 125.7 and JPE 212.7 versus EMPMP's 146.5/250.4, lowering APE by 10.3% and JPE by 15%.
• On ETH-UCY the wide model ties the best average ADE (0.18) and FDE (0.32) among compared methods; on SDD the deep model improves FDE by 32%.
• The deep model trains at 928 samples/s and tests at 3673 samples/s on MOCAP-UMPM versus EMPMP's 812 and 2041 on one RTX A6000.

Surprising findings

• Without any explicit interaction module, the model reaches strong multi-person results by treating individuals independently.
• Multi-modal (K=6) prediction improves the wide model's APE by 13.8% over its deterministic K=1 setting, while EMPMP's APE barely changes.

Limitations

• Joint training across all datasets shows a clear performance degradation, especially on AMASS, relative to specialized models.
• No explicit interaction modeling between agents in multi-person scenes.
• Deterministic K-proposal approach only competitive, not better than generative baselines, on diversity metrics (MMADE/MMFDE).

Experimental setup

Datasets: Human3.6M AMASS ETH-UCY SDD MOCAP-UMPM 3DPW

Benchmarks: Human3.6M AMASS ETH-UCY SDD MOCAP-UMPM 3DPW

Metrics: ADE FDE APE JPE MMADE MMFDE

Baselines: DivSamp BeLFusion CoMusion Motionmap SkeletonDiff SLD MID GP-Graph TUTR SingularTrajectory TrajCLIP NMRF

Builds on: DETR BeLFusion T2P EMPMP NMRF

Closest papers by meaning

• Social-Transmotion: Promptable Human Trajectory Prediction (ICLR 2024 · similarity 0.88)
• Expressive Forecasting of 3D Whole-Body Human Motions (AAAI 2024 · similarity 0.88)
• MotionMap: Representing Multimodality in Human Pose Forecasting (CVPR 2025 · similarity 0.88)
• Motion Diversification Networks (CVPR 2024 · similarity 0.87)
• FutureHuman3D: Forecasting Complex Long-Term 3D Human Behavior from Video Observations (CVPR 2024 · similarity 0.87)
• HumMUSS: Human Motion Understanding using State Space Models (CVPR 2024 · similarity 0.87)