Spaces:
Sleeping
Sleeping
Commit History
-
Android: ship pre-trained — bundle templates + lexicon from signs/
Gradle seedAssets copies signs/{lexicon,lsf_signs}.json into assets at build time (single source of truth); MainActivity seeds the on-device store on first launch. APK verified building (67 MB, 143-sample model bundled). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>de48e34 -
Add 2nd Elix video per word (143 samples) + one-command run.sh
- elix_train.py: --skip/--count, re-encode dedupe; 62 words now have 2 samples (143 total across 72 glosses) for better DTW generalisation. - The Python web app + CLI already load signs/lsf_signs.json, so they're pre-trained out of the box. Add run.sh: venv bootstrap + serve in one command. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>3b55737 -
Pre-train 72 LSF signs from Elix dictionary videos
Add tools/elix_train.py: fetches each catalogue word's reference sign video from dico.elix-lsf.fr and trains a DTW template via the existing pipeline. 72/81 words now ship pre-trained in signs/lsf_signs.json (the 9 misses are multi-word phrases). Polite (rate-limited, UA-identified; robots.txt permits crawling). Templates are abstract landmark data for personal use — the videos themselves are not stored. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>54effa2
-
Ignore local .claude/ tooling dir
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>c3e5941 -
Android: commit Gradle 8.7 wrapper; APK verified building
Built locally: Gradle 8.7 + AGP 8.5.2 + JDK 17 + SDK 34 -> app-debug.apk (64 MB, model + MediaPipe native libs bundled). Commit the wrapper so ./gradlew works everywhere; CI now uses it. README updated with the verified build steps. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>564f335 -
Add GitHub Actions workflow to build the Android APK in the cloud
Compiles app-debug.apk on push (no local Android Studio/SDK needed): JDK 17 + Gradle 8.7, fetches the MediaPipe model, uploads the APK as an artifact. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>5ab31b4 -
Add native Kotlin Android app (on-device, MediaPipe HolisticLandmarker)
Fully on-device LSF interpreter: - CameraX preview + front/back camera selector (selfie/main), front mirrored. - MediaPipe HolisticLandmarker on each frame (fingers/eyes/mouth/chest/arms). - Skeleton overlay (hand/pose lines, face dots). - DTW learn-by-example recognizer + segmentation + persistence, ported from the Python engine (identical thresholds); ~80-word LSF lexicon bundled. - Native dual-pixel/depth detection via Camera2 DEPTH_OUTPUT capability. Not compiler-verified (no Android toolchain here): open in Android Studio, run android/fetch_model.sh for the .task model, then build the APK. Excluded from the Docker/HF build context. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>fe3ddd9 -
Add front/back camera selector + per-camera mirroring
- Camera dropdown (enumerated devices) + flip button (facingMode user/environment). - Mirror only the selfie (front) camera; rear/main camera shown un-mirrored. - Forward-compatible with the planned on-device Android (WebView) build. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>e943395 -
Show face skeleton + speed up on CPU hosts
- Send face mesh in the landmark payload; browser draws face contours (eyes, eyebrows, lips, oval) — the face skeleton was previously omitted. - Lower inference resolution (PROC_WIDTH 480->320) and JPEG quality for much faster MediaPipe on CPU-only hosts (HF free tier). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>c4c0b6e -
coma error
e155725 -
Initial commit: LSF (French Sign Language) interpreter
Browser-camera Django web app + core pipeline: - MediaPipe Holistic landmarks (fingers/eyes/mouth/chest/arms) + JS skeleton overlay - DTW learn-by-example recognizer, ~80-word LSF lexicon with Elix video links - Client-side dual-pixel sensor detection (Pixel devices) - Video-clip importer for batch "pre-training" - Dockerised + Hugging Face Space ready (gunicorn + whitenoise) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>5029d29