DynaPURLS: Dynamic Refinement of Part-aware Representations for Skeleton-based Zero-Shot Action Recognition
Pioneering AI in Human Action Recognition
DynaPURLS introduces a novel framework for zero-shot skeleton-based action recognition, dynamically refining part-aware representations at inference time to achieve state-of-the-art accuracy.
Quantifiable Impact of DynaPURLS
Our innovative approach delivers significant performance gains across challenging benchmarks, demonstrating its potential for real-world enterprise applications.
0
ZSL Accuracy (NTU120)
0
GZSL Harmonic Mean (NTU120)
0
Improvement over SOTA (GZSL)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
DynaPURLS integrates multi-granularity semantic representation, adaptive partitioning, and dynamic test-time refinement. This unified framework establishes robust visual-semantic correspondences and adapts to unseen classes at inference.
Enterprise Process Flow
Generate Hierarchical Text Descriptions (LLM)
→
Adaptive Partitioning for Visual Features
→
Dynamic Test-Time Refinement
→
Confidence-Guided Optimization
→
Class-Balanced Memory Bank
→
Enhanced ZS-SAR Performance
9.3%
Average Classification Performance Improvement Post-adaptation (Figure 2)
The core of DynaPURLS lies in its adaptive mechanisms: a novel adaptive partitioning module and a dynamic test-time feature refinement, enabling robust knowledge transfer to unseen classes.
| Strategy | ZSL Acc. (NTU60 55/5) | GZSL H-Mean (NTU60 55/5) |
|---|---|---|
| Global (Original) | 64.69% | 35.46% |
| Global (GPT-3) | 78.50% | 33.47% |
| Static Partitioning | 76.46% | 33.03% |
| Adaptive Partitioning (DynaPURLS) | 79.23% | 40.99% |
Note: Adaptive Partitioning with dynamic refinement significantly outperforms static methods, especially in GZSL, demonstrating its ability to bridge the semantic-visual gap. |
||
Real-world Application: Enhanced Security Monitoring
Imagine a security system that can detect previously unseen anomalous behaviors. DynaPURLS's ability to recognize novel actions without prior training allows for the immediate identification of suspicious activities, significantly reducing response times and improving overall security posture.
Key Benefits:
- Early detection of emerging threats.
- Reduced false positives from known actions.
- Adaptability to evolving threat landscapes.
Extensive experiments on NTU RGB+D 60/120 and PKU-MMD demonstrate DynaPURLS's superior performance over prior art in both ZSL and GZSL settings.
| Method | ZSL (NTU60 55/5) | GZSL H (NTU60 55/5) | ZSL (NTU120 110/10) | GZSL H (NTU120 110/10) |
|---|---|---|---|---|
| Neuron [42] | 86.90% | 73.80% | 71.50% | 63.30% |
| SCoPLe [41] | 84.10% | 71.94% | 74.53% | 62.27% |
| PURLS [31] | 79.22% | 60.35% | 71.95% | 72.00% |
| DynaPURLS (Ours) | 88.52% | 69.95% | 89.06% | 81.49% |
Note: DynaPURLS consistently achieves state-of-the-art performance, with significant improvements over prior methods, especially on challenging GZSL scenarios. |
||||
Calculate Your Potential AI ROI
Estimate the potential annual savings and reclaimed hours for your enterprise by leveraging DynaPURLS's advanced action recognition capabilities.
Estimated Annual Savings
$0
Hours Reclaimed Annually
0
Your Path to Advanced HAR with DynaPURLS
A structured approach ensures seamless integration and maximum impact.
Phase 1: Needs Assessment & Data Preparation
Identify key action recognition challenges, prepare existing skeleton datasets, and define project scope.
Phase 2: Model Customization & Training
Adapt DynaPURLS backbone to your specific domain, fine-tune LLM prompts for nuanced action descriptions, and initial training on seen classes.
Phase 3: Deployment & Dynamic Refinement Integration
Deploy the pre-trained model and integrate the dynamic refinement module for real-time adaptation to novel actions.
Phase 4: Continuous Optimization & Scaling
Monitor performance, refine confidence thresholds, and expand to new action categories as your operational needs evolve.
Unlock the Future of Human Action Recognition
Ready to transform your enterprise's capabilities with state-of-the-art zero-shot skeleton-based action recognition?
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat