Lightcella Face Models (ncnn)
ncnn-format conversions of three permissively-licensed face models, for on-device use on low-power hardware (Raspberry Pi 4 / NAS). Format conversion only — no retraining. The detector and refiner were validated against their ONNX reference in onnxruntime.
| model | files | size | role | upstream | license |
|---|---|---|---|---|---|
| YuNet | detector/yunet_2023mar.ncnn.{param,bin} |
0.22 MB | face detection | OpenCV Zoo face_detection_yunet_2023mar |
MIT (© 2020 Shiqi Yu) |
| Face Mesh | refiner/mediapipe_facemesh_478.ncnn.{param,bin} |
4.8 MB | 478 face landmarks | MediaPipe face_landmarks_detector (face_landmarker.task, 2023-05-03) |
Apache-2.0 |
| GhostFaceNet | embedder/ghostfacenet_w1.3_s2.ncnn.{param,bin} |
8.1 MB | face embedding (512-d) | HamadYA/GhostFaceNets W1.3 S2 ArcFace | MIT (© 2022 HamadYA) |
Each subdirectory carries its model and its own LICENSE. Repo license is other because the
files carry two MIT + one Apache-2.0 (map in LICENSE.md).
I/O
| model | input | output |
|---|---|---|
| YuNet | RGB image, 640×640 | detection heads (cls/obj/bbox/kps) at strides 8/16/32 |
| Face Mesh | RGB face crop, 256×256 | 478 3D landmarks + presence score |
| GhostFaceNet | aligned RGB face, 112×112 | 512-d embedding |
Blob names follow the pnnx conversion (in0 / out0…). ncnn param/bin format.
Provenance
Converted with tf2onnx / onnxsim / pnnx and checked against onnxruntime. The exact upstream
source artifacts, SHA-256-pinned, are archived at
lightcella/face-pipeline-sources.
Attribution
Credit to the upstream authors — YuNet (Shiqi Yu, OpenCV Zoo), MediaPipe Face Mesh (Google),
GhostFaceNets (HamadYA). This repository redistributes ncnn-converted weights under their
respective upstream licenses; see LICENSE.md.