kyereboatengcaleb/palm-ccnet-onnx

Palm-print encoder (CCNet) — ONNX

ONNX export of the CCNet palm-print feature extractor for contactless palm recognition. Used by the face+palm verification backbone: it turns a normalised palm ROI into an L2-normalised embedding matched by cosine similarity.

Files

file	precision	size	use
palm_ccnet.onnx	fp32	~257 MB	servers (default fetch)
palm_ccnet_fp16.onnx	fp16	~128 MB	mobile / size-constrained (~lossless)
hand_landmarker.task	—	~8 MB	MediaPipe Hands ROI detector (Google)

Input / output contract

• Input: float32 NCHW [1, 1, 128, 128] — grayscale, pixel values in [0, 1].
• Output: [1, 2048] L2-normalised embedding. Match with cosine; calibrate the threshold on your data (the backbone does this adaptively).

Provenance & attribution

Exported (feature extractor only, not the close-set classifier) from the official CCNet pretrained Tongji checkpoint:

• CCNet — Yang et al., Comprehensive Competition Mechanism in Palmprint Recognition, IEEE TIFS 2023. Code: https://github.com/Zi-YuanYang/CCNet
• Trained on the Tongji contactless palmprint dataset (respect its terms).

This is a domain-pretrained model; for best accuracy on your own captures, fine-tune and re-export. You are responsible for license/dataset compliance in your use.