Spaces:
Sleeping
Sleeping
|
Download presentation/visuals/README.md from ofraij123/BridgeLinkASL: direct link, hf CLI and curl.
- Browser
- Download file 1.42 kB
-
https://huggingface.co/spaces/ofraij123/BridgeLinkASL/resolve/main/presentation/visuals/README.md
- Command line
-
hf download hf://spaces/ofraij123/BridgeLinkASL/presentation/visuals/README.md
-
curl -L -o README.md https://huggingface.co/spaces/ofraij123/BridgeLinkASL/resolve/main/presentation/visuals/README.md
1.42 kB
A newer version of the Gradio SDK is available: 6.32.0
Presentation Visuals
Generated slide-ready visuals for BridgeLink ASL.
Files
bridgelink_pipeline.svg: methodology / system diagramproject_scope_board.svg: what the team trained versus what was reusedcnn_vs_vlm_comparison.svg: final comparison numbers for the presentationmetrics-summary.json: source values used to render the comparison slidehow2sign_dataset_constraint.(png|svg): explains why the full 31k How2Sign clips cannot be used as sentence classeshow2sign_subset_benchmark.(png|svg): Top-12, Top-25, and Top-25 normalized benchmark growth and accuracyhow2sign_top25_class_distribution.(png|svg): class imbalance view for the normalized repeated-sentence subsethow2sign_top25_experiment_comparison.(png|svg): comparison of the Top-25 baseline, normalized checkpoint, continued training, and balanced-weight experimentshow2sign_plot_metrics.json: source values used by the matplotlib How2Sign plots
Current values
- Evaluation set: 36 clips
- Unique classes: 22
- CNN top-1: 25.0%
- CNN top-5: 58.3%
- Qwen rerank: 25.0%
- VLM wrapper failures: 0
Regenerate
python scripts\generate_presentation_visuals.py
python scripts\generate_how2sign_presentation_plots.py
Current How2Sign 3D CNN values
- Top-12 test accuracy: 14.7%
- Top-25 test accuracy: 20.8%
- Top-25 normalized test accuracy: 27.8%
- Top-25 normalized classes: 21