Appearance embedding

OSNet-AIN x1.0 (MSMT17)

Embed a pedestrian crop into a 512-dim vector for tracking across frames.

MIT

Explore model

About this model

TorchReID OSNet with Instance Adaptive Normalization, trained on MSMT17. Input is a 256x128 pedestrian crop; output is a 512-dim L2-normalized appearance embedding. Cosine similarity between two embeddings scores whether they are the same person.

About 2.2M parameters, 1 MB graph plus 8.6 MB external weights. CPU warm runs land at 20 ms per crop, leaving budget for the detector and tracker in a live pipeline.

What you can build

Maintain track IDs across frames in a multi-camera retail counter.
Re-identify a person who left and re-entered the camera view.
Score whether two detections in different frames match by appearance.
Drive a DeepSort-style tracker that survives bbox drift and brief occlusion.

Variants & downloads

Choose the build for your app

Model files are being prepared.

Explore the model’s examples, source links, and integration notes while download variants are added.

Example inference

A multi-person scene from the test pair, frame 1. — Input

Side-by-side composite with matched-identity boxes drawn across the two frames. — Output

Performance

Results depend on the device, runtime, and test conditions. Each measurement keeps its own context.

M1 Max, macOS 15, ONNX Runtime 1.17

See original catalog hardware notes · CPU

19.9 ms

Warm inference

Catalog reference · variant unspecifiedCold start (ms): 21.23

Reference result from the existing catalog; conversion variant and measurement date were not recorded.

M1 Max, macOS 15, ONNX Runtime 1.17

See original catalog hardware notes · CoreML

19.9 ms

Warm inference

Catalog reference · variant unspecifiedCold start (ms): 20.46

Reference result from the existing catalog; conversion variant and measurement date were not recorded.

RTX 2060, Ubuntu 24.04, ONNX Runtime 1.24.3

See original catalog hardware notes · CUDA

5.1 ms

Warm inference

Catalog reference · variant unspecified

Reference result from the existing catalog; conversion variant and measurement date were not recorded.

Integration examples

Vision SDK integration

cpp

#include "imaged/sdk.hpp"
#include "imaged/families/reid/reid_family.hpp"

imaged::SDK::AI ai;
auto& config = ai.getOptions();
config.set("reid.model", "weights/osnet_ain_x1_0_msmt17.onnx");
config.set("reid.inputHeight", 256);
config.set("reid.inputWidth", 128);

imaged::SDK::ReIDFamily family;
family.load(config);

imaged::SDK::Image personCrop;
personCrop.load("crop.jpg");

auto result = std::get<imaged::SDK::EmbeddingResult>(family.run({&personCrop}));
auto& features = result.features;  // 512-dim unit vector