Touch Eyes Data Analysis

This repository contains the Android/iOS touch data collection code, labeled raw dataset archives, trained CNN model assets, and full ablation reports for thumb/index light/heavy touch classification.

Layout

• android/: Android collection app, reusable aispect-core module, training and ablation scripts.
• ZHIFAeyes/ and ZHIFAeyes.xcodeproj/: current iOS workspace.
• data/lemon/: labeled raw JSON archive for the Xiaomi device/person dataset.
• data/tao/: labeled raw JSON archive for the OnePlus device/person dataset.
• data/manifests/: full four-class manifest and dataset summary used for training.
• models/lemon_tao_full4_top5/: top 5 four-class CNN models from the full lemon + tao ablation.
• models/reports/: full CNN ablation CSV/JSON reports.
• docs/: GitHub Pages dashboard files.

Dataset Policy

The raw expanded folders lemon/, tao/, and lemon_raw_before_thumb_mark/ are local working copies and are intentionally ignored by Git. Upload uses compressed archives under data/ to avoid committing tens of thousands of small JSON files.

The full four-class training manifest contains 30,214 labeled JSON samples:

• lemon: 17,872 samples
• tao: 12,342 samples
• Classes: thumb_light, thumb_heavy, index_light, index_heavy
• No quality filtering
• No downsampling
• No confidence channel in the best deployed model

Current Deployed Model Order

The Android catalog is ordered with the latest full lemon + tao four-class models first:

1. lemon_tao_full4_enhanced_patch21_no_confidence
2. lemon_tao_full4_enhanced_patch22
3. lemon_tao_full4_old13_multichannel
4. lemon_tao_full4_old13_gaussian_summary
5. lemon_tao_full4_old13_raw_radius

These assets are also embedded in android/aispect-core/src/main/assets/aispect_impact_models/ so the debug APK can switch models on-device.