Confidence-Aware DSTA-SLR

Confidence-Aware DSTA-SLR is a research-oriented extension of DSTA-SLR for skeleton-based sign language recognition. This release focuses on making pose-confidence signals usable throughout the training and evaluation pipeline, with support for confidence-aware modeling, reliability-aware consistency learning, robustness analysis, and reproducible experiment workflows.

The public repository root is intentionally lightweight. It serves as the entry point to the main codebase, experiment scripts, and release-facing documentation, while large artifacts stay out of version control.

This repository is centered on one practical question:

How much do skeleton confidence signals help sign language recognition, and how can they be used in a controlled, reproducible way?

To answer that, the codebase extends the original DSTA-SLR pipeline with:

• confidence-aware input handling and confidence perturbation controls
• model-side confidence encoding and aggregation changes
• reliability-aware consistency losses for training
• experiment runners for ablation, robustness, and repeat studies
• utilities for stream fusion and result reporting

• a clean public entry point for the project release
• the main research code under DSTA-SLR/
• runnable experiment workflows for baseline, confidence-aware, and robustness settings
• documentation for scripts, run order, and reproduction boundaries

• Main codebase: DSTA-SLR/
• Main usage guide: DSTA-SLR/readme.md
• Script index: DSTA-SLR/scripts/README.md
• Experiment runbook: DSTA-SLR/scripts/EXPERIMENT_RUNBOOK.md
• License: DSTA-SLR/LICENSE

If you are opening this repository for the first time, the fastest path is:

1. Read DSTA-SLR/readme.md for the project overview and setup.
2. Check DSTA-SLR/scripts/README.md to understand the script layout.
3. Use DSTA-SLR/scripts/EXPERIMENT_RUNBOOK.md to follow the intended experiment order.

The public release includes the code and documentation needed to understand and reproduce the confidence-aware DSTA-SLR workflows:

• training and evaluation code
• model modifications
• experiment orchestration scripts
• fusion and reporting utilities
• public-facing documentation

• Data should be placed locally under DSTA-SLR/data/.
• Experiment outputs should be kept under DSTA-SLR/work_dir/ or other ignored local directories.
• Large checkpoints are better distributed through GitHub Releases or external storage than through normal Git history.

This release builds on the original DSTA-SLR line of work and related open-source sign language recognition projects referenced from the main codebase documentation.