Research map / Graph learning / GNN architectures
GCNext: Towards the Unity of Graph Convolutions for Human Motion Prediction
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 UniGC, a universal graph convolution unifying spatial, temporal, channel graph convolutions via adjacency masking and sub-adjacency tying, and GCNext, a layer-wise and sample-wise dynamic selector over GC types for human motion prediction.
Problem
Human motion prediction with GCNs relies on manually chosen graph convolution types; the best type varies across layers and samples, and the full 6D global adjacency is parameter-heavy and inefficient.
Approach
UniGC re-conceptualizes existing graph convolutions (spatial, temporal, spatial-temporal, channel, G3D, STS, etc.) as special cases of one 6D global adjacency via masking and sub-adjacency tying; GCNext stacks layers with several GC candidates and a Gumbel-Softmax selector that picks one per sample, plus dimension-preserving update and layer normalization.
Key results
- On Human3.6M at 1000ms, GCNext achieves 108.7mm average MPJPE versus 109.4mm for siMLPe and 114.0mm for LTD.
- Under the averaged protocol on Human3.6M, GCNext reaches 64.7mm at 1000ms versus 71.6mm for MotionMixer.
- GCNext uses 29.1M train FLOPs (14.5M inference) versus 133.3M for LTD and 1449.6M for MSR-GCN, claiming up to 9x lower computational cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GCNext instead of siMLPe under the per-frame evaluation protocol | Human3.6M · Human3.6M, per-frame protocol, 1000ms | MPJPE (mm): 109.4 -> 108.7 | Helped |
| Use GCNext instead of MotionMixer under the average-frame evaluation protocol | Human3.6M · Human3.6M, average-frame protocol, 1000ms | MPJPE (mm): 71.6 -> 64.7 | Helped |
| Refine an existing GCN with GCNext | LTD on Human3.6M · Human3.6M, 1000ms | MPJPE (mm): 114.0 -> 110.5 | Helped |
| Refine an existing GCN with GCNext | MSR-GCN on Human3.6M · Human3.6M, 1000ms | MPJPE (mm): 112.9 -> 110.8 | Helped |
| Refine an existing GCN with GCNext | PGBIG on Human3.6M · Human3.6M, 1000ms | MPJPE (mm): 109.7 -> 108.8 | Helped |
| Replace dynamic selection with one fixed graph-convolution type (ablation) | GCNext · Human3.6M, 1000ms | MPJPE (mm): 108.7 vs 109.9–116.0 | Hurt |
| Replace dynamic selection with averaging all graph-convolution outputs (ablation) | GCNext · Human3.6M, 1000ms | MPJPE (mm): 108.7 vs 109.6 | Hurt |
| Replace dynamic selection with summing all graph-convolution outputs (ablation) | GCNext · Human3.6M, 1000ms | MPJPE (mm): 108.7 vs 110.9 | Hurt |
| Use the four options spatial-temporal, spatial-channel, spatial, and channel graph convolutions (ablation) | GCNext selector · Human3.6M, 1000ms | MPJPE (mm): 109.5 -> 108.7 | Helped |
| Use 48 layers rather than 24 layers (ablation) | GCNext · Human3.6M, 1000ms | MPJPE (mm): 109.8 -> 108.7 | Helped |
| Use 48 layers rather than 96 layers (ablation) | GCNext · Human3.6M, 1000ms | MPJPE (mm): 110.2 -> 108.7 | Helped |
| Use the unbiased dynamic policy instead of random or graph-convolution-biased policies (ablation) | GCNext · Human3.6M, 1000ms | MPJPE (mm) | Helped |
Surprising findings
- Using all graph convolution candidates by averaging or summing outputs performs worse than the dynamic selector, and the full 6D global adjacency (G) performs worst among static variants at 116.0mm at 1000ms.
Limitations
- Memory usage (3.1 GB) is higher than some baselines such as LTD (2.2 GB) and HisRep (2.3 GB).
- Gains over the strongest baseline siMLPe are modest on several settings (e.g., 3DPW and AMASS at some horizons).
Experimental setup
Datasets: Human3.6M AMASS 3DPW
Benchmarks: Human3.6M AMASS-BMLrub 3DPW
Metrics: MPJPE
Baselines: LTD HisRep MSR-GCN PGBIG siMLPe STSGCN GAGCN MotionMixer
Builds on: G3D LTD siMLPe Gumbel Softmax
Closest papers by meaning
- Spatio-Temporal Gating-Adjacency GCN for Human Motion Prediction (CVPR 2022 · similarity 0.89)
- L2-GCN: Layer-Wise and Learned Efficient Training of Graph Convolutional Networks (CVPR 2020 · similarity 0.89)
- Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning (AAAI 2018 · similarity 0.88)
- Sequential Graph Convolutional Network for Active Learning (CVPR 2021 · similarity 0.88)
- Are Spatial-Temporal Graph Convolution Networks for Human Action Recognition Over-Parameterized? (CVPR 2025 · similarity 0.88)
- Motion Graph Unleashed: A Novel Approach to Video Prediction (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.