Elements 2.0

meta_overlay

Overlays the metadata on the video frame to visualize the inference results.

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - video/x-raw
  SRC template: src	Availability: Always Capabilities: - video/x-raw

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  async-handling	The bin will handle Asynchronous state changes Default:False
  message-forward	Forwards all children messages Default:False
  preprocess	Pre-processing element Default: None
  process	Main processing element Default: None
  postprocess	Post-processing element Default: None
  aggregate	(Optional) Element to aggregate preprocess/process/postprocess result and original frame Default: None
  postaggregate	(Optional) Element inserted after aggregation element Default: None
  preprocess-queue-size	Size of queue (in number buffers) before pre-processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  process-queue-size	Size of queue (in number buffers) before processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  postprocess-queue-size	Size of queue (in number buffers) before post-processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  aggregate-queue-size	Size of queue (in number buffers) for original frames between ‘tee’ and aggregate element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  postaggregate-queue-size	Size of queue (in number buffers) between aggregate and post-aggregate elements. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  device	Target device for meta_overlaying Default: <enum CPU device on system memory of type MetaOverlayDevice>

object_classify

Performs object classification. Accepts the ROI or full frame as an input and outputs classification results with metadata.

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - video/x-raw
  SRC template: src	Availability: Always Capabilities: - video/x-raw

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  async-handling	The bin will handle Asynchronous state changes Default:False
  message-forward	Forwards all children messages Default:False
  preprocess	Pre-processing element Default: None
  process	Main processing element Default: None
  postprocess	Post-processing element Default: None
  aggregate	(Optional) Element to aggregate preprocess/process/postprocess result and original frame Default: None
  postaggregate	(Optional) Element inserted after aggregation element Default: None
  preprocess-queue-size	Size of queue (in number buffers) before pre-processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  process-queue-size	Size of queue (in number buffers) before processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  postprocess-queue-size	Size of queue (in number buffers) before post-processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  aggregate-queue-size	Size of queue (in number buffers) for original frames between ‘tee’ and aggregate element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  postaggregate-queue-size	Size of queue (in number buffers) between aggregate and post-aggregate elements. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  model	Path to inference model network file Default: “”
  ie-config	Comma separated list of KEY=VALUE parameters for inference configuration Default: “”
  device	Target device for inference. Please see inference backend documentation (ex, OpenVINO™ Toolkit) for list of supported devices. Default: CPU
  model-instance-id	Identifier for sharing resources between inference elements of the same type. Elements with the instance-id will share model and other properties. If not specified, a unique identifier will be generated. Default: “”
  nireq	Maximum number of inference requests running in parallel Default: 0
  batch-size	Number of frames batched together for a single inference. If the batch-size is 0, then it will be set by default to be optimal for the device. Not all models support batching. Use model optimizer to ensure that the model has batching support. Default: 0
  model-proc	Path to JSON file with parameters describing how to build pre-process and post-process sub-pipelines Default: “”
  pre-process-backend	Preprocessing backend type Default: <enum Automatic of type VideoInferenceBackend>
  inference-interval	Run inference for every Nth frame Default: 1
  roi-inference-interval	Determines how often to run inference on each ROI object. Only valid if each ROI object has unique object id (requires object tracking after object detection) Default: 1
  inference-region	Region on which inference will be performed - full-frame or on each ROI (region of interest)bounding-box area Default: <enum Perform inference for full frame of type VideoInferenceRegion>
  object-class	Run inference only on Region-Of-Interest with specified object class Default: “”
  labels	Path to file containing model’s output layer labels or comma separated list of KEY=VALUE pairs where KEY is name of output layer and VALUE is path to labels file. If provided, labels from model-proc won’t be loaded Default: “”
  labels-file	Path to file containing model’s output layer labels. If provided, labels from model-proc won’t be loaded Default: “”
  attach-tensor-data	If true, metadata will contain both post-processing results and raw tensor data. If false, metadata will contain post-processing results only. Default: True
  threshold	Threshold for detection results. Only regions of interest with confidence values above the threshold will be added to the frame. Zero means default (auto-selected) threshold Default: 0.0*
  scale-method	Scale method to use in pre-preprocessing before inference Default: <enum Default of type VideoInferenceScaleMethod>
  repeat-metadata	If true and inference-interval > 1, metadata with last inference results will be attached to frames if inference skipped. If true and roi-inference-interval > 1, it requires object-id for each roi, so requires object tracking element inserted before this element. Default:False
  reclassify-interval	Determines how often to reclassify tracked objects. Only valid when used in conjunction with gvatrack. The following values are acceptable: 0 - Do not reclassify tracked objects 1 - Always reclassify tracked objects 2:N - Tracked objects will be reclassified every N frames. Note the inference-interval is applied before determining if an object is to be reclassified (i.e. classification only occurs at a multiple of the inference interval) Default: 1

object_detect

Performs inference-based object detection

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - video/x-raw
  SRC template: src	Availability: Always Capabilities: - video/x-raw

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  async-handling	The bin will handle Asynchronous state changes Default:False
  message-forward	Forwards all children messages Default:False
  preprocess	Pre-processing element Default: None
  process	Main processing element Default: None
  postprocess	Post-processing element Default: None
  aggregate	(Optional) Element to aggregate preprocess/process/postprocess result and original frame Default: None
  postaggregate	(Optional) Element inserted after aggregation element Default: None
  preprocess-queue-size	Size of queue (in number buffers) before pre-processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  process-queue-size	Size of queue (in number buffers) before processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  postprocess-queue-size	Size of queue (in number buffers) before post-processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  aggregate-queue-size	Size of queue (in number buffers) for original frames between ‘tee’ and aggregate element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  postaggregate-queue-size	Size of queue (in number buffers) between aggregate and post-aggregate elements. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  model	Path to inference model network file Default: “”
  ie-config	Comma separated list of KEY=VALUE parameters for inference configuration Default: “”
  device	Target device for inference. Please see inference backend documentation (ex, OpenVINO™ Toolkit) for list of supported devices. Default: CPU
  model-instance-id	Identifier for sharing resources between inference elements of the same type. Elements with the instance-id will share model and other properties. If not specified, a unique identifier will be generated. Default: “”
  nireq	Maximum number of inference requests running in parallel Default: 0
  batch-size	Number of frames batched together for a single inference. If the batch-size is 0, then it will be set by default to be optimal for the device. Not all models support batching. Use model optimizer to ensure that the model has batching support. Default: 0
  model-proc	Path to JSON file with parameters describing how to build pre-process and post-process sub-pipelines Default: “”
  pre-process-backend	Preprocessing backend type Default: <enum Automatic of type VideoInferenceBackend>
  inference-interval	Run inference for every Nth frame Default: 1
  roi-inference-interval	Determines how often to run inference on each ROI object. Only valid if each ROI object has unique object id (requires object tracking after object detection) Default: 1
  inference-region	Region on which inference will be performed - full-frame or on each ROI (region of interest)bounding-box area Default: <enum Perform inference for full frame of type VideoInferenceRegion>
  object-class	Run inference only on Region-Of-Interest with specified object class Default: “”
  labels	Path to file containing model’s output layer labels or comma separated list of KEY=VALUE pairs where KEY is name of output layer and VALUE is path to labels file. If provided, labels from model-proc won’t be loaded Default: “”
  labels-file	Path to file containing model’s output layer labels. If provided, labels from model-proc won’t be loaded Default: “”
  attach-tensor-data	If true, metadata will contain both post-processing results and raw tensor data. If false, metadata will contain post-processing results only. Default: True
  threshold	Threshold for detection results. Only regions of interest with confidence values above the threshold will be added to the frame. Zero means default (auto-selected) threshold Default: 0.0*
  scale-method	Scale method to use in pre-preprocessing before inference Default: <enum Default of type VideoInferenceScaleMethod>
  repeat-metadata	If true and inference-interval > 1, metadata with last inference results will be attached to frames if inference skipped. If true and roi-inference-interval > 1, it requires object-id for each roi, so requires object tracking element inserted before this element. Default:False

object_track

Assigns unique ID to detected objects

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - video/x-raw
  SRC template: src	Availability: Always Capabilities: - video/x-raw

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  async-handling	The bin will handle Asynchronous state changes Default:False
  message-forward	Forwards all children messages Default:False
  preprocess	Pre-processing element Default: None
  process	Main processing element Default: None
  postprocess	Post-processing element Default: None
  aggregate	(Optional) Element to aggregate preprocess/process/postprocess result and original frame Default: None
  postaggregate	(Optional) Element inserted after aggregation element Default: None
  preprocess-queue-size	Size of queue (in number buffers) before pre-processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  process-queue-size	Size of queue (in number buffers) before processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  postprocess-queue-size	Size of queue (in number buffers) before post-processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  aggregate-queue-size	Size of queue (in number buffers) for original frames between ‘tee’ and aggregate element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  postaggregate-queue-size	Size of queue (in number buffers) between aggregate and post-aggregate elements. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  model	Path to inference model network file Default: “”
  ie-config	Comma separated list of KEY=VALUE parameters for inference configuration Default: “”
  device	Target device for inference. Please see inference backend documentation (ex, OpenVINO™ Toolkit) for list of supported devices. Default: CPU
  model-instance-id	Identifier for sharing resources between inference elements of the same type. Elements with the instance-id will share model and other properties. If not specified, a unique identifier will be generated. Default: “”
  nireq	Maximum number of inference requests running in parallel Default: 0
  batch-size	Number of frames batched together for a single inference. If the batch-size is 0, then it will be set by default to be optimal for the device. Not all models support batching. Use model optimizer to ensure that the model has batching support. Default: 0
  model-proc	Path to JSON file with parameters describing how to build pre-process and post-process sub-pipelines Default: “”
  pre-process-backend	Preprocessing backend type Default: <enum Automatic of type VideoInferenceBackend>
  inference-interval	Run inference for every Nth frame Default: 1
  roi-inference-interval	Determines how often to run inference on each ROI object. Only valid if each ROI object has unique object id (requires object tracking after object detection) Default: 1
  inference-region	Region on which inference will be performed - full-frame or on each ROI (region of interest)bounding-box area Default: <enum Perform inference for full frame of type VideoInferenceRegion>
  object-class	Run inference only on Region-Of-Interest with specified object class Default: “”
  labels	Path to file containing model’s output layer labels or comma separated list of KEY=VALUE pairs where KEY is name of output layer and VALUE is path to labels file. If provided, labels from model-proc won’t be loaded Default: “”
  labels-file	Path to file containing model’s output layer labels. If provided, labels from model-proc won’t be loaded Default: “”
  attach-tensor-data	If true, metadata will contain both post-processing results and raw tensor data. If false, metadata will contain post-processing results only. Default: True
  threshold	Threshold for detection results. Only regions of interest with confidence values above the threshold will be added to the frame. Zero means default (auto-selected) threshold Default: 0.0*
  scale-method	Scale method to use in pre-preprocessing before inference Default: <enum Default of type VideoInferenceScaleMethod>
  repeat-metadata	If true and inference-interval > 1, metadata with last inference results will be attached to frames if inference skipped. If true and roi-inference-interval > 1, it requires object-id for each roi, so requires object tracking element inserted before this element. Default:False
  generate-objects	If true, generate objects (according to previous trajectory) if not detected on current frame Default: True
  adjust-objects	If true, adjust object position for more smooth trajectory Default: True
  tracking-per-class	If true, object association takes into account object class Default:False
  spatial-feature	Spatial feature used by object tracking algorithm Default: <enum Spatial feature not used (only temporal features used, such as object shape and trajectory) of type SpatialFeatureType>
  spatial-feature-distance	Method to calculate distance between two spatial features Default: <enum Spatial feature not used of type SpatialFeatureDistanceType>
  tracking-type	DEPRECATED - use other properties according to the following mapping: zero-term-imageless: generate-objects=false adjust-objects=false spatial-feature=none zero- term: generate-objects=false adjust-objects=false spatial-feature=sliced-histogram short-term-imageless: gene rate-objects=true adjust-objects=false spatial-feature=none short-term: generate-objects=true adjust-objects=false spatial-feature=sliced-histogram Default: “”

processbin

Bin element for processing pipelines using branching: tee name=t t. ! <preprocess> ! <process> ! <postprocess> ! <aggregate> t. ! aggregate

• Capabilities

  SINK template: sink	Availability: Always
  SRC template: src	Availability: Always

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  async-handling	The bin will handle Asynchronous state changes Default:False
  message-forward	Forwards all children messages Default:False
  preprocess	Pre-processing element Default: None
  process	Main processing element Default: None
  postprocess	Post-processing element Default: None
  aggregate	(Optional) Element to aggregate preprocess/process/postprocess result and original frame Default: None
  postaggregate	(Optional) Element inserted after aggregation element Default: None
  preprocess-queue-size	Size of queue (in number buffers) before pre-processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  process-queue-size	Size of queue (in number buffers) before processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  postprocess-queue-size	Size of queue (in number buffers) before post-processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  aggregate-queue-size	Size of queue (in number buffers) for original frames between ‘tee’ and aggregate element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  postaggregate-queue-size	Size of queue (in number buffers) between aggregate and post-aggregate elements. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0

video_inference

Runs Deep Learning inference on any model with RGB-like input

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - video/x-raw
  SRC template: src	Availability: Always Capabilities: - video/x-raw

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  async-handling	The bin will handle Asynchronous state changes Default:False
  message-forward	Forwards all children messages Default:False
  preprocess	Pre-processing element Default: None
  process	Main processing element Default: None
  postprocess	Post-processing element Default: None
  aggregate	(Optional) Element to aggregate preprocess/process/postprocess result and original frame Default: None
  postaggregate	(Optional) Element inserted after aggregation element Default: None
  preprocess-queue-size	Size of queue (in number buffers) before pre-processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  process-queue-size	Size of queue (in number buffers) before processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  postprocess-queue-size	Size of queue (in number buffers) before post-processing element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  aggregate-queue-size	Size of queue (in number buffers) for original frames between ‘tee’ and aggregate element. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  postaggregate-queue-size	Size of queue (in number buffers) between aggregate and post-aggregate elements. Special values: -1 means no queue element, 0 means queue of unlimited size Default: 0
  model	Path to inference model network file Default: “”
  ie-config	Comma separated list of KEY=VALUE parameters for inference configuration Default: “”
  device	Target device for inference. Please see inference backend documentation (ex, OpenVINO™ Toolkit) for list of supported devices. Default: CPU
  model-instance-id	Identifier for sharing resources between inference elements of the same type. Elements with the instance-id will share model and other properties. If not specified, a unique identifier will be generated. Default: “”
  nireq	Maximum number of inference requests running in parallel Default: 0
  batch-size	Number of frames batched together for a single inference. If the batch-size is 0, then it will be set by default to be optimal for the device. Not all models support batching. Use model optimizer to ensure that the model has batching support. Default: 0
  model-proc	Path to JSON file with parameters describing how to build pre-process and post-process sub-pipelines Default: “”
  pre-process-backend	Preprocessing backend type Default: <enum Automatic of type VideoInferenceBackend>
  inference-interval	Run inference for every Nth frame Default: 1
  roi-inference-interval	Determines how often to run inference on each ROI object. Only valid if each ROI object has unique object id (requires object tracking after object detection) Default: 1
  inference-region	Region on which inference will be performed - full-frame or on each ROI (region of interest)bounding-box area Default: <enum Perform inference for full frame of type VideoInferenceRegion>
  object-class	Run inference only on Region-Of-Interest with specified object class Default: “”
  labels	Path to file containing model’s output layer labels or comma separated list of KEY=VALUE pairs where KEY is name of output layer and VALUE is path to labels file. If provided, labels from model-proc won’t be loaded Default: “”
  labels-file	Path to file containing model’s output layer labels. If provided, labels from model-proc won’t be loaded Default: “”
  attach-tensor-data	If true, metadata will contain both post-processing results and raw tensor data. If false, metadata will contain post-processing results only. Default: True
  threshold	Threshold for detection results. Only regions of interest with confidence values above the threshold will be added to the frame. Zero means default (auto-selected) threshold Default: 0.0*
  scale-method	Scale method to use in pre-preprocessing before inference Default: <enum Default of type VideoInferenceScaleMethod>
  repeat-metadata	If true and inference-interval > 1, metadata with last inference results will be attached to frames if inference skipped. If true and roi-inference-interval > 1, it requires object-id for each roi, so requires object tracking element inserted before this element. Default:False

batch_create

Accumulate multiple buffers into single buffer with multiple GstMemory

• Capabilities

  SINK template: sink	Availability: Always
  SRC template: src	Availability: Always

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  batch-size	Number of frames to batch together Default: 1

batch_split

Split input tensor (remove batch dimension from tensor shape)

• Capabilities

  SINK template: sink	Availability: Always
  SRC template: src	Availability: Always

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False

capsrelax

Pass data without modification, relaxes formats

• Capabilities

  SINK template: sink	Availability: Always
  SRC template: src	Availability: Always

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False

gvadrop

Pass / drop custom number of frames in pipeline

• Capabilities

  SINK template: sink	Availability: Always
  SRC template: src	Availability: Always

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  pass-frames	Number of frames to pass along the pipeline. Default: 1
  drop-frames	Number of frames to drop. Default: 0
  mode	Mode defines what to do with dropped frames Default:

meta_aggregate

Muxes video streams with tensor’s ROI into single stream

• Capabilities

  SINK template: meta%u	Availability: On request Capabilities: - video/x-raw
  SINK template: tensor%u	Availability: On request Capabilities: - other/tensors
  SINK template: sink	- Availability: Always
  SRC template: src	- Availability: Always

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  latency	Additional latency in live mode to allow upstream to take longer to produce buffers for the current position (in nanoseconds) Default: 0
  min-upstream-latency	When sources with a higher latency are expected to be plugged in dynamically after the aggregator has started playing, this allows overriding the minimum latency reported by the initial source(s). This is only taken into account when larger than the actually reported minimum latency. (nanoseconds) Default: 0
  start-time-selection	Decides which start time is output Default: <enum GST_AGGREGATOR_START_TIME_SELECTION_ZERO of type GstAggregatorStartTimeSelection>
  start-time	Start time to use if start-time-selection=set Default: 18446744073709551615
  emit-signals	Send signals Default:False
  attach-tensor-data	If true, additionally copies tensor data into metadata Default: True

meta_smooth

smooth metadata

• Capabilities

  SINK template: sink	Availability: Always
  SRC template: src	Availability: Always

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False

roi_split

Split buffer with multiple GstVideoRegionOfInterestMeta into multiple buffers

• Capabilities

  SINK template: sink	Availability: Always
  SRC template: src	Availability: Always

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  object-class	Filter ROI list by object class(es) (comma separated list if multiple). Output only ROIs with specified object class(es) Default: “”

video_frames_buffer

Buffer and optionally repeat compressed video frames

• Capabilities

  SINK template: sink	Availability: Always
  SRC template: src	Availability: Always

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  num-input-frames	Number input frames to buffer Default: 0
  num-output-frames	Max number output frames in ‘loop’ mode Default: 0

rate_adjust

Adjust frame rate. Output frame rate is input rate multiplied by (numerator/denominator)

• Capabilities

  SINK template: sink	Availability: Always
  SRC template: src	Availability: Always

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  ratio	Frame rate ratio - output frame rate is input rate multiplied by specified ratio. Current limitation: ratio <= 1 Default: None

tensor_convert

Convert (zero-copy if possible) between video/audio and tensors media type

• Capabilities

  SINK template: sink	Availability: Always Capabilities: video/x-raw format: RGB video/x-raw format: BGR video/x-raw format: RGBA video/x-raw format: BGRA video/x-raw format: RGBP video/x-raw format: BGRP
  SRC template: src	Availability: Always Capabilities: other/tensors num_tensors: 1 types: uint8

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False

tensor_histogram

Calculates histogram on tensors of UInt8 data type and NHWC layout

• Capabilities

  SINK template: sink	Availability: Always Capabilities: other/tensors num_tensors: 1 types: uint8
  SRC template: src	Availability: Always Capabilities: other/tensors num_tensors: 1 types: float32

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  width	Input tensor width, assuming tensor in NHWC or NCHW layout Default: 64
  height	Input tensor height, assuming tensor in NHWC or NCHW layout Default: 64
  num-slices-x	Number slices along X-axis Default: 1
  num-slices-y	Number slices along Y-axis Default: 1
  num-bins	Number bins in histogram calculation. Example, for 3-channel tensor (RGB image), output histogram size is equal to (num_bin^3 * num_slices_x * num_slices_y) Default: 8
  batch-size	Batch size Default: 1
  device	CPU or GPU or GPU.0, GPU.1, .. Default: “”

tensor_postproc_add_params

Post-processing to only add properties/parameters to metadata

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - other/tensors
  SRC template: src	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  attribute-name	Name for metadata created and attached by this element Default: attribute
  format	Format description Default: “”

tensor_postproc_detection

Post-processing of object detection inference to extract bounding box coordinates, confidence, label, mask

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - other/tensors
  SRC template: src	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  labels	Array of object classes Default: None
  labels-file	Path to .txt file containing object classes (one per line) Default: “”
  threshold	Detection threshold - only objects with confidence values above the threshold will be added to the frame Default: 0.5
  box-index	Index of layer containing bounding box data Default: -1
  confidence-index	Index of layer containing confidence data Default: -1
  label-index	Index of layer containing label data Default: -1
  imageid-index	Index of layer containing imageid data Default: -1
  mask-index	Index of layer containing mask data Default: -1
  box-offset	Offset inside layer containing bounding box data Default: -1
  confidence-offset	Offset inside layer containing confidence data Default: -1
  label-offset	Offset inside layer containing label data Default: -1
  imageid-offset	Offset inside layer containing imageid data Default: -1

tensor_postproc_label

Post-processing of classification inference to extract object classes

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - other/tensors
  SRC template: src	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  method	Method used to post-process tensor data Default:
  labels	Array of object classes Default: None
  labels-file	Path to .txt file containing object classes (one per line) Default: “”
  attribute-name	Name for metadata created and attached by this element Default: “”
  layer-name	Name of output layer to process (in case of multiple output tensors) Default: “”
  threshold	Threshold for confidence values Default: 0.0*
  compound-threshold	Threshold for compound method Default: 0.5

tensor_postproc_text

Post-processing to convert tensor data into text

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - other/tensors
  SRC template: src	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  text-scale	Scale tensor values before converting to text Default: 1.0
  text-precision	Precision for floating-point to text conversion Default: 0
  attribute-name	Name for metadata created and attached by this element Default: “”
  layer-name	Name of output layer to process (in case of multiple output tensors) Default: “”

tensor_postproc_yolo

Post-processing of YOLO models to extract bounding box list

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - other/tensors
  SRC template: src	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  version	Yolo’s version number. Supported only from 3 to 5 Default: 0
  labels	Array of object classes Default: None
  labels-file	Path to .txt file containing object classes (one per line) Default: “”
  threshold	Detection threshold - only objects with confidence value above the threshold will be added to the frame Default: 0.5
  anchors	Anchor values array Default: None
  masks	Masks values array (1 dimension) Default: None
  iou-threshold	IntersectionOverUnion threshold Default: 0.5
  do-cls-softmax	If true, perform softmax Default: True
  output-sigmoid-activation	output_sigmoid_activation Default: True
  cells-number	Number of cells. Use if number of cells along x and y axes is the same (0 = autodetection) Default: 0
  cells-number-x	Number of cells along x-axis Default: 0
  cells-number-y	Number of cells along y-axis Default: 0
  bbox-number-on-cell	Number of bounding boxes that can be predicted per cell (0 = autodetection) Default: 0
  classes	Number of classes Default: 0
  nms	Apply Non-Maximum Suppression (NMS) filter to bounding boxes Default: True

tensor_sliding_window

Sliding aggregation of input tensors

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - other/tensors
  SRC template: src	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False

openvino_tensor_inference

Inference on OpenVINO™ toolkit backend

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - other/tensors - other/tensors
  SRC template: src	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  model	Path to model file in OpenVINO™ toolkit or ONNX format Default: “”
  device	Target device for inference. Please see OpenVINO™ toolkit documentation for list of supported devices. Default: CPU
  config	Comma separated list of KEY=VALUE parameters for Inference Engine configuration Default: “”
  batch-size	Batch size Default: 1
  buffer-pool-size	Output buffer pool size (functionally same as OpenVINO™ toolkit nireq parameter) Default: 16
  shared-instance-id	Identifier for sharing backend instance between multiple elements, for example in elements processing multiple inputs Default: “”

openvino_video_inference

Inference on OpenVINO™ toolkit backend

• Capabilities

  SINK template: sink	Availability: Always Capabilities: video/x-raw format: NV12
  SRC template: src	Availability: Always Capabilities: other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  model	Path to model file in OpenVINO™ toolkit or ONNX format Default: “”
  device	Target device for inference. Please see OpenVINO™ toolkit documentation for list of supported devices. Default: CPU
  config	Comma separated list of KEY=VALUE parameters for Inference Engine configuration Default: “”
  batch-size	Batch size Default: 1
  buffer-pool-size	Output buffer pool size (functionally same as OpenVINO™ toolkit nireq parameter) Default: 16
  shared-instance-id	Identifier for sharing backend instance between multiple elements, for example in elements processing multiple inputs Default: “”

opencv_cropscale

Fused video crop and scale on OpenCV backend. Crop operation supports GstVideoCropMeta if attached to input buffer

• Capabilities

  SINK template: sink	Availability: Always Capabilities: video/x-raw format: RGB video/x-raw format: BGR video/x-raw format: RGBA video/x-raw format: BGRA
  SRC template: src	Availability: Always Capabilities: video/x-raw format: RGB video/x-raw format: BGR video/x-raw format: RGBA video/x-raw format: BGRA

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  add-borders	Add borders if necessary to keep the aspect ratio Default:False

opencv_find_contours

Find contour points of given mask using opencv

• Capabilities

  SINK template: sink	Availability: Always
  SRC template: src	Availability: Always

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  mask-metadata-name	Name of metadata containing segmentation mask Default: mask
  contour-metadata-name	Name of metadata created by this element to store contour(s) Default: contour
  threshold	Mask threshold - only mask pixels with confidence values above the threshold will be used for finding contours Default: 0.5

opencv_meta_overlay

Visualize inference results using OpenCV

• Capabilities

  SINK template: sink	Availability: Always Capabilities: video/x-raw format: BGRA video/x-raw format: RGBA video/x-raw format: BGRA video/x-raw format: RGBA
  SRC template: src	Availability: Always Capabilities: video/x-raw format: BGRA video/x-raw format: RGBA video/x-raw format: BGRA video/x-raw format: RGBA

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  lines-thickness	Thickness of lines and rectangles Default: 2
  font-thickness	Font thickness Default: 1
  font-scale	Font scale Default: 1.0
  attach-label-mask	Attach label mask as metadata, image not changed Default:False

opencv_object_association

Assigns unique ID to ROI objects based on objects trajectory and (optionally) feature vector obtained from ROI metadata

• Capabilities

  SINK template: sink	Availability: Always
  SRC template: src	Availability: Always

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  generate-objects	If true, generate objects (according to previous trajectory) if not detected on current frame Default: True
  adjust-objects	If true, adjust object position for more smooth trajectory Default: True
  tracking-per-class	If true, object association takes into account object class Default:False
  spatial-feature-metadata-name	Name of metadata containing spatial feature Default: spatial-feature
  spatial-feature-distance	Method to calculate distance between two spatial features Default: <enum bhattacharyya of type spatial-feature-distance>
  shape-feature-weight	Weighting factor for shape-based feature Default: 0.75
  trajectory-feature-weight	Weighting factor for trajectory-based feature Default: 0.5
  spatial-feature-weight	Weighting factor for spatial feature Default: 0.25
  min-region-ratio-in-boundary	> Min region ratio in image > boundary Default: 0.75

opencv_remove_background

Remove background using mask

• Capabilities

  SINK template: sink	Availability: Always Capabilities: video/x-raw format: RGB video/x-raw format: BGR video/x-raw format: RGBA video/x-raw format: BGRA
  SRC template: src	Availability: Always Capabilities: video/x-raw format: RGB video/x-raw format: BGR video/x-raw format: RGBA video/x-raw format: BGRA

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  mask-metadata-name	Name of metadata containing segmentation mask Default: mask
  threshold	Mask threshold - only mask pixels with confidence values above the threshold will be used for setting transparency Default: 0.5

opencv_tensor_normalize

Convert U8 tensor to F32 tensor with normalization

• Capabilities

  SINK template: sink	Availability: Always Capabilities: other/tensors num_tensors: 1 types: uint8
  SRC template: src	Availability: Always Capabilities: other/tensors num_tensors: 1 types: float32

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  range	Normalization range MIN, MAX. Example: <0,1> Default: None
  mean	Mean values per channel. Example: <0.485,0.456,0.406> Default: None
  std	Standard deviation values per channel. Example: <0.229,0.224,0.225> Default: None

opencv_warp_affine

Rotation using cv::warpAffine

• Capabilities

  SINK template: sink	Availability: Always Capabilities: video/x-raw format: RGB video/x-raw format: BGR video/x-raw format: RGBA video/x-raw format: BGRA
  SRC template: src	Availability: Always Capabilities: video/x-raw format: RGB video/x-raw format: BGR video/x-raw format: RGBA video/x-raw format: BGRA

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  angle	Angle by which the picture is rotated (in radians) Default: 0.0*
  sync	Wait for OpenCL kernel completion (if running on GPU via cv::UMat) Default:False

tensor_postproc_human_pose

Post-processing to extract key points from human pose estimation model output

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - other/tensors
  SRC template: src	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  point-names	Array of key point names Default: None
  point-connections	Array of point connections {name-A0, name-B0, name-A1, name-B1, …} Default: None

vaapi_batch_proc

Batched pre-processing with VAAPI memory as input and output

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - video/x-raw
  SRC template: src	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  add-borders	Add borders if necessary to keep the aspect ratio Default:False
  output-format	Image format for output frames: BGR or RGB or GRAY Default: BGR
  shared-instance-id	Identifier for sharing backend instance between multiple elements, for example in elements processing multiple inputs Default: “”

vaapi_sync

Synchronize VAAPI surfaces (call vaSyncSurface)

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - video/x-raw
  SRC template: src	Availability: Always Capabilities: - video/x-raw

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  timeout	Synchronization timeout (seconds) Default: 10.0

opencl_tensor_normalize

Convert U8 tensor to U8 or F32 tensor with normalization

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - other/tensors
  SRC template: src	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  shared-instance-id	Identifier for sharing backend instance between multiple elements, for example in elements processing multiple inputs Default: “”

vaapi_to_opencl

Convert memory:VASurface to memory:OpenCL

• Capabilities

  SINK template: sink	Availability: Always Capabilities: - video/x-raw - other/tensors
  SRC template: src	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False

sycl_meta_overlay

Visualize inference results using DPC++/SYCL backend

• Capabilities

  SINK template: sink	Availability: Always Capabilities: video/x-raw format: BGRA video/x-raw format: RGBA
  SRC template: src	Availability: Always Capabilities: video/x-raw format: BGRA video/x-raw format: RGBA

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  lines-thickness	Thickness of lines and rectangles Default: 2

sycl_tensor_histogram

Calculates histogram on tensors of UInt8 data type and NHWC layout

• Capabilities

  SINK template: sink	Availability: Always Capabilities: other/tensors num_tensors: 1 types: uint8 other/tensors num_tensors: 1 types: uint8
  SRC template: src	Availability: Always Capabilities: other/tensors num_tensors: 1 types: float32

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  width	Input tensor width, assuming tensor in NHWC or NCHW layout Default: 64
  height	Input tensor height, assuming tensor in NHWC or NCHW layout Default: 64
  num-slices-x	Number slices along X-axis Default: 1
  num-slices-y	Number slices along Y-axis Default: 1
  num-bins	Number bins in histogram calculation. Example, for 3-channel tensor (RGB image), output histogram size is equal to (num_bin^3 * num_slices_x * num_slices_y) Default: 8
  batch-size	Batch size Default: 1
  device	CPU or GPU or GPU.0, GPU.1, .. Default: “”

inference_openvino

OpenVINO™ toolkit inference element

• Capabilities

  SRC template: src	Availability: Always Capabilities: - other/tensors
  SINK template: sink	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  device	Inference device Default: CPU
  model	OpenVINO™ toolkit model path Default: “”
  nireq	Number inference requests Default: 0

pytorch_tensor_inference

PyTorch inference element

• Capabilities

  SRC template: src	Availability: Always Capabilities: - other/tensors
  SINK template: sink	Availability: Always Capabilities: - other/tensors

• Properties

  Name	Description
  name	The name of the object Default: None
  parent	The parent of the object Default: None
  qos	Handle Quality-of-Service events Default:False
  device	Inference device Default: cpu
  model	The full module name of the PyTorch model to be imported from torchvision or model path. Ex. ’torchvision.models.resnet50’ or ’/path/to/model.pth’ Default: “”
  model-weights	PyTorch model weights path. If model-weights is empty, the default weights will be used Default: “”