person_vit_clip_reid

CLIP-ReID person re-identification model exported to ONNX. Reconstructs the visual tower of CLIP ViT-B/16 with the projection head learned during person-reID fine-tuning.

• Architecture: CLIP ViT-B/16 visual encoder
• Training set: Market-1501
• Source checkpoint: Syliz517/CLIP-ReID — vanilla ViT-CLIP-ReID for Market-1501
• Input: image tensor NCHW, float32, ImageNet normalization
• Output: L2-normalized embedding, dim 512
• File: person_vit_clip_reid.onnx (self-contained, no external-data files)

Usage

import onnxruntime as ort
sess = ort.InferenceSession("person_vit_clip_reid.onnx",
                            providers=["CUDAExecutionProvider", "CPUExecutionProvider"])
embedding = sess.run(None, {"input": chw_image_batch})[0]

Source license

Inherits CLIP-ReID's MIT license.