person_reid_siglip2_naflex_512d

person_reid_siglip2_naflex_512d

Image-to-image person re-identification. Crops in, 512-d unit-norm vectors out; compare with a dot product (cosine similarity).

Crops are consumed at their natural aspect ratio — no fixed 256×128 resize.

Benchmarks

mAP / rank-1 under the standard re-ID protocol (gallery entries sharing both identity and camera with the query are dropped as junk; queries left with no true match are skipped). Best per row in bold.

Notes:

• The real-camera surveillance set is an in-house benchmark of fixed-camera captures — the closest proxy to deployment conditions. On its 200 queries the top-1 match is wrong 5 times for this model and 22 times for CLIP-ReID.
• The mixed-domain gallery pools three public test splits into one gallery, so most distractors are out-of-domain relative to any given query. It penalises models that collapse away from their home domain.
• ENTIRe-ID is unseen by every model in the table.
• FastReID BoT-R50 and NVIDIA ReID Transformer are the publicly released weights. The FastReID column uses the Market-trained release; its MSMT17-trained release scores 0.2197 / 0.2197 on ENTIRe-ID.
• Table figures are measured at the model's native 768-d width. The published 512-d projection retains 99.8% of embedding variance and measures within noise: 0.9180 / 0.9700 on the real-camera set.

Files

ONNX interface — three inputs, one output:

pixel_values          (B, 256, 768)  float32   patch sequence, padded
pixel_attention_mask  (B, 256)       int64     1 = real patch, 0 = padding
spatial_shapes        (B, 2)         int64     per-image patch grid (rows, cols)
→ embedding           (B, 512)       float32   L2-normalized in-graph

Preprocessing: RGB, natural aspect ratio, patch-16 tokenisation to at most 256 patches, mean/std 0.5, BILINEAR. Guards: minimum side 32 px, aspect ratio capped at 5:1.