Strict liveness metrics, reported honestly.
TinyLiveness currently ships one FP32 ONNX release model. The website and backend use the stricter APCER 1% policy by default.
APCER 1% release variant
The following metrics are based on a corrected, balanced 400-image test slice containing 200 live and 200 spoof samples. The slice was sampled from a larger 4,000-image live/spoof dataset, with results averaged across 10 separate evaluation windows.
- ROC AUC
- 0.999325
- PR AUC
- 0.999332
- Accuracy
- 98.25%
- APCER
- 1.00%
- BPCER
- 2.50%
- ACER
- 1.75%
- BPCER100
- 3.00%
- Threshold
- 0.990000
Model size and runtime
The model is intended for CPU-friendly API deployments and compact mobile pipelines. The backend route loads FP32 ONNX with CPU execution.
- FP32 ONNX size
- 15.296 MB
- Synthetic CPU latency
- 5.619 ms/image
- Input tensor
- Nx3x224x224
- Normalization
- ImageNet