| --- |
| library_name: onnx |
| license: other |
| license_name: upstream-model-and-dataset-terms |
| license_link: https://huggingface.co/Shusek00/ray-local-models/blob/main/README.md#provenance-and-terms |
| tags: |
| - face-recognition |
| - face-detection |
| - onnx |
| - fp16 |
| - mobile |
| base_model: |
| - minchul/cvlface_adaface_ir50_webface4m |
| --- |
| |
| # Ray Local model exports |
|
|
| Deployment exports for the Ray Local Suvio plugin, evaluated on Apple M1 Max and a |
| Qualcomm SM8850 Android device. These are converted pretrained models, not new training runs. |
|
|
| | File | Bytes | Role | |
| | --- | ---: | --- | |
| | `tinyfacematch-fp16.onnx` | 6,948,481 | Default recognizer, 128 dimensions | |
| | `adaface-ir50-webface4m-fp16.onnx` | 87,227,974 | Optional larger recognizer, 512 dimensions | |
| | `yunet-2026may-dynamic-fp32.onnx` | 229,738 | Shared face detector and five landmarks | |
|
|
| The base pair is 7,178,219 bytes. All three current files together are 94,406,193 bytes. |
| The earlier `yunet-2023mar-640-fp32.onnx` remains available for reproducing previous |
| benchmarks; current plugin packages include only the dynamic detector. |
| `manifest.json` records exact SHA-256 digests and tensor contracts. Applications should |
| pin an immutable repository commit and verify both length and digest before opening a file. |
|
|
| ## Tensor contracts |
|
|
| Both recognizers take `input`: float32 `[1, 3, 112, 112]`, RGB, NCHW, after five-point |
| ArcFace alignment. Float16 weights and internal computation retain float32 public I/O. |
|
|
| - TinyFaceMatch: `(pixel - 127.5) / 128.0`; `embedding`: float32 `[1, 128]`, L2 normalized. |
| - AdaFace: `(pixel - 127.5) / 127.5`; `embedding`: float32 `[1, 512]`. L2 normalize the output. |
| This CVLFace export uses RGB; do not substitute the BGR convention of other AdaFace exports. |
| - YuNet: `input`: float32 `[1, 3, height, width]`, BGR pixels in `[0, 255]`. |
| Pad each spatial dimension to a multiple of 32. Ray limits the longest source edge to |
| 640 without upscaling, then adds zero padding at the right and bottom. Decode `cls_*`, |
| `obj_*`, `bbox_*`, `kps_*` at strides 8, 16, 32 using the actual padded dimensions. |
| A 640×360 frame uses 640×384. The dynamic model shares the 2023 model's learned weights. |
|
|
| The detector is still required. A larger embedding model does not replace face detection |
| or correct inaccurate landmarks. Embeddings from different recognizers are incompatible. |
|
|
| ## Measurements |
|
|
| Aggregate public benchmark evaluation on prealigned crops, canonical ten-fold held-out |
| threshold selection, no flip augmentation: |
|
|
| | Export | CFP-FP, 7,000 pairs | CPLFW, 6,000 pairs | |
| | --- | ---: | ---: | |
| | TinyFaceMatch FP16 | 96.1143% | 90.8667% | |
| | AdaFace FP16 | 98.9429% | 93.9167% | |
|
|
| Warm batch-one recognition latency with synthetic inputs: TinyFaceMatch / AdaFace |
| 1.251 / 3.724 ms on M1 Max CoreML, and 0.920 / 4.689 ms on Qualcomm QNN HTP. |
| Android strict QNN profiles recorded accelerator execution with CPU graph fallback disabled. |
| These are standalone model measurements; application integration and session loading add cost. |
|
|
| Android incremental warm PSS with YuNet on CPU: approximately 149 / 443 MB. |
| These earlier combined-memory measurements used the fixed 640×640 detector. |
| Dynamic detector geometry was separately checked on synthetic inputs on M1 Max; |
| recognition inputs remain static 112×112. QNN EP requires static shapes, so the dynamic |
| detector uses CPU in the current host while supported recognizers can use QNN HTP. |
| This is process memory, not a measurement of all NPU memory. File size is not runtime memory. |
| No identity labels, user photos, per-pair scores or embeddings are distributed here. |
|
|
| ## Provenance and terms |
|
|
| This repository does not grant new rights to upstream pretrained weights or training data. |
|
|
| - **AdaFace**: [CVLFace AdaFace IR50 WebFace4M](https://huggingface.co/minchul/cvlface_adaface_ir50_webface4m), |
| by Minchul Kim and collaborators. The upstream model card requires following the training |
| dataset's license. See [AdaFace (CVPR 2022)](https://arxiv.org/abs/2204.00964) and |
| [CVLFace](https://github.com/mk-minchul/CVLface). ONNX export at opset 17; static batch one; |
| FP16 conversion with ONNX Runtime 1.29.0 and ONNX 1.22.0. |
| - **TinyFaceMatch**: [yuvrajraina/tinyfacematch](https://github.com/yuvrajraina/tinyfacematch), |
| `tinyfacematch-128-pretrained.onnx`, upstream SHA-256 |
| `6d8588c1dc1f91fab930be355d33d4b6be0b74d70c46ae0f9c65d89be2865aa4`. |
| The repository code is MIT, but the model metadata and export script identify |
| InsightFace `buffalo_s/w600k_mbf.onnx` as its pretrained base, followed by PCA projection. |
| [InsightFace's pretrained model terms](https://github.com/deepinsight/insightface/tree/master/model_zoo) |
| restrict the provided weights to non-commercial research. Do not treat the wrapper's MIT |
| license as commercial clearance for these derivative weights. |
| - **YuNet**: [OpenCV Zoo](https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet), |
| `face_detection_yunet_2026may.onnx`, revision `47534e27c9851bb1128ccc0102f1145e27f23f98`, |
| MIT model directory license. The current file is byte-identical to the official dynamic |
| FP32 export; it is not a retrained or larger detector. The legacy static file is retained |
| separately for reproducibility. |
|
|
| For any intended distribution or use, the applicable upstream model and dataset terms remain in force. |
| This export repository makes no independent claim of commercial permission. |
|
|
