Spaces:
Sleeping
Sleeping
| # Project Execution Plan | |
| This plan treats BridgeLink ASL as one integrated project rather than a role | |
| handoff. | |
| ## Goal | |
| Complete a Hugging Face Space proof of concept backed by reproducible WLASL | |
| experiments: | |
| ```text | |
| WLASL-100 training | |
| -> MediaPipe landmark extraction | |
| -> temporal CNN baseline | |
| -> WLASL-25 demo checkpoint | |
| -> Qwen2.5-VL reranking comparison | |
| -> metrics, charts, Space demo, CVPR-style report | |
| ``` | |
| ## Required Work | |
| 1. Train the landmark CNN on WLASL-100. | |
| 2. Export the smaller WLASL-25 demo checkpoint for presentation use. | |
| 3. Build the WLASL-25 hybrid eval manifest with CNN top-5 candidates. | |
| 4. Run local Qwen2.5-VL reranking on the same clip set. | |
| 5. Generate `results/` artifacts with accuracy, precision, recall, F1, confusion matrix, and comparison rows. | |
| 6. Use Hugging Face Space as the live demo surface for the CNN app. | |
| 7. Write the CVPR-style report from the docs and generated results. | |
| ## Done Criteria | |
| - Space demo runs with webcam and uploaded clips. | |
| - Dataset visuals reflect the WLASL evaluation subset. | |
| - Experiments show CNN versus VLM metrics honestly. | |
| - Repo includes scripts to regenerate metrics and presentation assets. | |
| - Final report includes dataset, methodology, hyperparameters, experiment setup, metrics, graphs, and limitations. | |