Extract Appearance Node

E

AI/ML/Tracking

Crops every detection from the frame and runs a re-identification model on the crops in batches. The built-in models are Intel OpenVINO person re-identification models (Apache-2.0): the chosen one is downloaded into Cache Dir on first use, checked against its SHA-256 and read from there afterwards. With Model set to custom, connect your own session from Load ONNX and set Normalization to match its export: imagenet for torchreid/OSNet exports, raw for FastReID onnx_export.py exports, which normalize inside the model. Each detection becomes an observation with an L2-normalized appearance embedding, keeping its box, camera, session, tracker id, track id and state (track-only fields such as hits and velocity are not carried). Lost tracks from Track Detections (Include Lost) carry a predicted box, so they are passed through with their existing embedding instead of being cropped. Feed the result into Track Detections for appearance-aware tracking or into Associate Entities to recognize the same object across cameras. Boxes that are not finite or smaller than Min Box Size are left out.

tracking_extract_appearanceonnx

Inputs15

Outputs3

Security exposure8/10

Packageonnx

Ratings

Scores range from 0 to 10. Higher values mean more impact, exposure, or operational weight.

SecurityAttack surface and exposure impact.

8/10High

PrivacyPotential sensitivity of processed data.

5/10Medium

PerformanceRuntime or resource pressure.

6/10Medium

GovernancePolicy, audit, or compliance impact.

5/10Medium

ReliabilityOperational stability considerations.

8/10High

CostExternal or compute cost impact.

9/10High

Input Pins

15

Input

Execution

exec_in

Initiate Execution

Model

String

model

person-openvino-0270 (6 MB, default) or person-openvino-0265 (9.6 MB): Intel OpenVINO person re-identification models, Apache-2.0, downloaded into Cache Dir on first use. custom uses the Custom Model pin.

Default person-openvino-0270

person-openvino-0270person-openvino-0265custom

Cache Dir

Struct

cache_dir

Folder the built-in model is downloaded to when missing and loaded from when present. Not used with a custom model.

SchemaFlowPathFlowPathFlowPath3 fields

pathstringrequired

store_refstringrequired

cache_store_refstring | null

Schema enforced

Custom Model

Struct

custom_model

Session from Load ONNX, used when Model is custom. Takes an RGB float tensor [N,3,H,W] or [N,H,W,3] and returns one embedding per crop, shaped [N,D] or [N,D,1,1].

SchemaNodeOnnxSessionNodeOnnxSessionNodeOnnxSession1 fields

session_refstringrequired

Cache ID for Session

Schema enforced

Image

Struct

image_in

Frame the detections were found in

SchemaNodeImageNodeImageNodeImage1 fields

image_refstringrequired

Schema enforced

Detections

Struct Array

detections

Boxes in image pixel coordinates, from Object Detection or tracks from Track Detections

SchemaBoundingBoxBoundingBoxBoundingBox7 fields

x1number:floatrequired

format float

y1number:floatrequired

format float

x2number:floatrequired

format float

y2number:floatrequired

format float

scorenumber:floatrequired

format float

class_idxinteger:int32required

format int32

class_namestring | null

Default []

Camera ID

String

camera_id

Camera stamped on every observation; empty keeps the incoming value

Session ID

String

session_id

Camera session stamped on every observation; empty keeps the incoming value

Timestamp (ms)

Integer

timestamp_ms

Frame capture time in Unix milliseconds; 0 keeps the incoming value, or uses the current time when there is none

Default 0

Normalization

String

normalization

Pixel scaling a custom model expects (the built-in models normalize their input themselves, so any value works for them): imagenet ((x/255 - mean) / std) for torchreid/OSNet exports, raw (0–255 unchanged) for FastReID onnx_export.py exports that normalize inside the model, zero_one (x/255) or minus_one_one (x/127.5 - 1)

Default imagenet

imagenetrawzero_oneminus_one_one

Input Width

Integer

input_width

Crop width in pixels, used only when the model's input width is dynamic

Default 128

Input Height

Integer

input_height

Crop height in pixels, used only when the model's input height is dynamic

Default 256

Padding

Float

padding

Fraction of the box width and height added on each side before cropping

Default 0

Range 0 to 1

Min Box Size

Float

min_box_size

Boxes narrower or shorter than this many pixels after padding and clipping to the image are skipped

Default 4

Batch Size

Integer

batch_size

Crops per inference call, used only when the model's batch dimension is dynamic

Default 16

Output Pins

3

Output

Execution

exec_out

Done with the Execution

Observations

Struct Array

observations

One observation per cropped detection with its new embedding, plus every lost track with its existing one; each carries its index in Detections

SchemaAppearanceObservationAppearanceObservationAppearanceObservation15 fields

x1number:floatrequired

format float

y1number:floatrequired

format float

x2number:floatrequired

format float

y2number:floatrequired

format float

scorenumber:floatrequired

format float

class_idxinteger:int32required

format int32

class_namestring | null

embeddingArray<number:float>

L2-normalized appearance embedding; empty when none was extracted

default []

itemsnumber:floatarray item

format float

camera_idstring

Camera the observation came from

default ""

session_idstring

Stream session of that camera

default ""

tracker_idstring

Tracker instance that issued `track_id`; track ids are only unique per tracker instance

default ""

track_idinteger | null

Local track id from Track Detections, if the observation belongs to a track

format uint64default nullmin 0

stateTrackState

State of the source track; plain detections are `tracked`

default "tracked"

variant 1constvariant

Matched to a detection in the latest frame

const "tracked"

variant 2constvariant

Not matched recently; kept alive for re-identification until it expires

const "lost"

timestamp_msinteger:int64

Frame capture time in Unix milliseconds

format int64default 0

detection_indexinteger | null

Index of the source element in the producing node's input array

format uint32default nullmin 0

Default []

Dimensions

Integer

dimensions

Embedding length of the model output; 0 when no detection was cropped

Node Info

Internal name

tracking_extract_appearance

Category

AI/ML/Tracking

Version

1