Human motion prediction repository with state-of-the-art models (2024-2025).
This repository has been significantly updated with cutting-edge techniques from 2024-2025 research:
- ✅ Graph Convolutional Networks (GCN) - Skeleton-aware predictions (+5-11% improvement)
- ✅ Temporal Convolutional Networks (TCN) - Efficient parallel training (+3-7% improvement)
- ✅ TCN-Transformer Hybrid - Best of both worlds (+7-12% improvement)
- ✅ Original RNN/Transformer - Updated with modern techniques
- ✅ Advanced Loss Functions - Velocity, acceleration, bone length constraints
- ✅ Uncertainty Estimation - Monte Carlo Dropout & ensemble methods
- ✅ Improved Training - Adam optimizer, learning rate scheduling, better metrics
- ✅ Comprehensive Metrics - MPJPE, PCK, calibration error
# Install dependencies
pip install torch numpy h5py matplotlib
# Train with Graph Convolutional Network (recommended)
cd human-motion-prediction-pytorch
python src/train_improved.py --model_type gcn --action walking --iterations 50000
# Train with Temporal Convolutional Network (fastest)
python src/train_improved.py --model_type tcn --action all --optimizer adam
# Original RNN model
python src/translate.py --action walking --seq_length_out 25 --iterations 10000- IMPROVEMENTS_2024.md - Detailed guide to all new features and models
- human-motion-prediction-pytorch/README.md - Original repository documentation
| Model | Speed | Accuracy | Memory | Use Case |
|---|---|---|---|---|
| GCN ⭐ | Fast | Best (+11%) | Medium | Best overall performance |
| TCN | Fastest | Great (+7%) | Low | Real-time applications |
| TCNFormer | Medium | Excellent (+12%) | Medium | Highest quality predictions |
| RNN (Original) | Slow | Baseline | Medium | Reproducibility |
Built on latest research from top conferences:
- 🏆 MST-GNN (IEEE TIP 2021) - Graph neural networks for motion
- 🏆 TCNFormer (2024) - Hybrid architecture
- 🏆 Uncertainty Estimation (RAL 2024) - Reliable forecasting
- 🏆 CoMusion (ECCV 2024) - Consistent stochastic prediction
- 🏆 Multi-Agent Forecasting (CVPR 2024) - Interaction modeling
See IMPROVEMENTS_2024.md for full details and citations.
MIT