MediaPipe Face Mesh V2 landmarks, converted to ONNX

This is Google's MediaPipe face-landmark network (face_landmarks_detector.tflite inside face_landmarker.task, float16 bundle v1), converted to ONNX for ONNX Runtime. The weights are Google's, under Apache 2.0; see the Face Mesh V2 model card. The model card places identification and life-critical decisions outside the model's intended use.

• Input input_12: float32 [N, 256, 256, 3], RGB, values in [0, 1], a face crop.
• Outputs: Identity [N, 1, 1, 1434] = 478 landmarks × (x, y, z) in input pixels; Identity_1 [N, 1, 1, 1] face-presence logit; Identity_2 [N, 1].
• Converted with tf2onnx (opset 17) from the TFLite file. On 5 random inputs the outputs match the TFLite interpreter within 0.0005.
• SHA-256 of face_landmarks.onnx: f38c3321ceffbc9e95103480ad38cc3f52e7e1bde2bcee7cd9355d0b9138ac0c

Source model: https://storage.googleapis.com/mediapipe-models/face_landmarker/face_landmarker/float16/1/face_landmarker.task (SHA-256 64184e229b263107bc2b804c6625db1341ff2bb731874b0bcc2fe6544e0bc9ff).

Changes from the original: format conversion only. No retraining or weight changes.