---
language:
  - fr
tags:
  - sign-language
  - pose-landmarks
  - mediapipe
  - transformer
  - african-sign-language
  - CASL
  - isolated-sign-recognition
license: c-uda
datasets:
  - luciayen/CASL-W60-Landmarks
metrics:
  - accuracy
model-index:
  - name: AfriSignEncoder Exp1 — CASL Baseline
    results:
      - task:
          type: video-classification
          name: Isolated Sign Word Recognition
        dataset:
          name: CASL-W60-Landmarks
          type: luciayen/CASL-W60-Landmarks
        metrics:
          - type: accuracy
            value: 0.7192
            name: Validation Accuracy
---

# AfriSignEncoder Exp1 — CASL Baseline (LandmarkTransformer)

Part of the **AfriSignEncoder** research project: a multilingual African sign language recognition benchmark.
This checkpoint is the **Experiment 1 single-language baseline** for Central African Sign Language (CASL).

## Model Description

A ViT-style transformer that treats T=64 MediaPipe Holistic landmark frames as a token sequence.

| Component | Value |
|---|---|
| Architecture | LandmarkTransformer (custom) |
| Input | (B, 64, 225) float32 — 75 keypoints × 3 coords per frame |
| Embedding dim | 256 |
| Attention heads | 8 |
| Encoder layers | 4 |
| Feed-forward dim | 1,024 |
| Positional encoding | Learned |
| Classification head | Linear 256 → 60 |
| Parameters | ~3.25 M |

## Dataset

**CASL-W60** — 60 word-level Central African Sign Language glosses.

| Split | Samples |
|---|---|
| Train | 3,667 |
| Test | 2,222 |

Source: Kaggle `mwakalucky/casl-w60` → parquet at `luciayen/CASL-W60-Landmarks`.
Landmark format: MediaPipe Holistic (pose 33 + left hand 21 + right hand 21) × xyz = 225D per frame.

## Training

| Setting | Value |
|---|---|
| Optimiser | AdamW (lr=3e-4, wd=1e-4) |
| LR schedule | OneCycleLR cosine |
| Max epochs | 60 |
| Batch size | 64 |
| Loss | CrossEntropy + label_smoothing=0.1 |
| Early stopping | patience=12 on val acc |
| Normalisation | Per-feature z-score (stats stored in checkpoint) |

## Results

| Metric | Value |
|---|---|
| Best validation accuracy | **71.92%** |
| Best checkpoint epoch | 35 |
| Final epoch (early stop) | 47 |
| Random-chance baseline | 1.67% |

## Checkpoint Contents

```python
import torch
ck = torch.load("pytorch_model.bin", map_location="cpu")
# Keys: epoch, val_acc, model (state_dict), l2i (label→index dict),
#        mean (tensor 225,), std (tensor 225,)
```

The `l2i` dict maps 60 CASL gloss strings to integer class indices.
`mean` and `std` are the per-feature normalisation statistics computed on the training set.

## Limitations

- Landmark-only (no RGB appearance) — sets a lower bound for CASL recognition.
- Train/test split is from the original dataset; signer independence has not been verified.
- 60 glosses is a small fraction of full CASL vocabulary.

## Citation / Project

AfriSignEncoder research project, CMU, 2026. GitHub: `africansl_encoder`.