SamDNX's picture
Initial commit: LSF (French Sign Language) interpreter
5029d29
|
Raw
History Blame Contribute Delete
1.14 kB
# Trained model hook (optional)
If you train a sequence classifier and drop it here, the recogniser uses it
automatically (in preference to the DTW templates).
Expected files:
| File | Purpose |
|-----------------------|----------------------------------------------------|
| `model.h5` | Keras model saved with `model.save("model.h5")` |
| `model.labels.json` | JSON list of class names, index-aligned to outputs |
**Input shape:** `(batch, 32, 153)` — i.e. each sample is a sign segment
resampled to `TEMPLATE_LEN = 32` frames of the *matching vector* (pose block +
both hands), `9*3 + 21*3*2 = 153` features per frame. This is exactly the array
the recogniser passes to `model.predict(...)` (see
`recognizer.SignRecognizer._classify_model`).
**Output:** a softmax over the classes listed in `model.labels.json`. Anything
below 0.5 confidence is ignored.
To build a dataset, record signs through the UI (they land in
`signs/lsf_signs.json` as `(32, 153)` sequences per gloss) and use those as
labelled training samples, augmenting with time/scale jitter.