GstInference Releases

GstInference
Introduction
Getting started
Getting the code Building the plugin Quick Starting Guide
Supported architectures
InceptionV1 InceptionV2 InceptionV3 InceptionV4 TinyYoloV2 TinyYoloV3 YoloV2 ResNetV1 MobileNetV2 MobileNetV2 + SSD AlexNet
Supported backends
TensorFlow Tensorflow-Lite Coral from Google Caffe TensorRT ONNXRT ONNXRT ACL ONNXRT OpenVINO
Metadata and Signals
GstClassificationMeta GstEmbeddingMeta GstDetectionMeta GstInferenceMeta Signals
Overlay Elements
Inference Overlay
Utils Elements
InferenceFilter InferenceCrop InferenceDebug InferenceBin
Legacy pipelines
PC NANO TX2 Xavier IMX8
Example pipelines
Example applications
Classification Detection Hand Pose Recognition
Benchmarks
Model Zoo
Project Status
Releases Roadmap
Contact Us

This page summarizes all GstInference and R2Inference releases. Check GitHub for more info!

v0.1.0

Introduced features:

• Video inference base class
• Support for the following architectures:
  • GoogLeNet (Inception v4)
  • Tiny Yolo v2
• Support for the following backends:
  • NCSDK
  • TensorFlow
• Support for the following platforms:
  • Intel Movidius Neural Compute Stick (version 1)
  • NVIDIA Jetson AGX Xavier
  • x86 systems
• Set backend from GStreamer
• Set backend properties from GStreamer
• Set model location from GStreamer
• Support play-stop-play transitions on GstInference elements

Known issues:

• NCSDK does not support multiple calls to the inference engine from the same thread. This causes transitions when using the NCSDK backend to fail after the second play.
• Changing the backend on a stoped pipeline will fail with a segmentation fault.
• When using the TensorFlow backend compiled with GPU support, the pipeline will sometimes fail to start.
• The plugins throw segmentation fault when using nvarguscamerasrc as input on Xavier.

v0.2.0

Introduced features:

• Stability improved.
• Inference on InceptionV4 improved.
• Metadata structure created to save the inference result.
• Metadata is attached to model buffer and bypass buffer (if available).
• Signal created to use the inference result at the application level.
• The inference result is not displayed by the inference element unless the log level is set to 6.
• Pads control improved, bypass pad is now optional.
• GoogleNet renamed to InceptionV4.
• TinyYolo renamed to TinyYoloV2.
• Jetson TX2 added to supported architectures.
• Examples section improved to smaller pipelines.
• Overlay plugins added:
  • GstClassificationOverlay for InceptionV4 metadata.
  • GstDetectionOverlay for TinyYoloV2 metadata.
• Example applications were added.
• The segmentation fault with NVarguscamerasrc is fixed.
• TensorFlow is able to run on CPU and GPU.

Known issues

• NCSDK does not support multiple calls to the inference engine from the same thread. This causes transitions when using the NCSDK backend to fail after the second play.
• Changing the backend on a stoped pipeline will fail with a segmentation fault.

v0.3.0

Introduced features:

• Fixed bug with gtk-doc compilation on GStreamer 1.8
• Support for the following architectures:
  • Inception v3
  • Inception v2
  • Inception v1

Known issues

• NCSDK does not support multiple calls to the inference engine from the same thread. This causes transitions when using the NCSDK backend to fail after the second play.
• Changing the backend on a stoped pipeline will fail with a segmentation fault.

v0.4.0

Introduced features:

• Support on Inception and TinyYOLO elements for the following image formats:
  • RGBx
  • RGBA
  • xRGB
  • ARGB
  • BGR
  • BGRx
  • BGRA
  • xBGR
  • ABGR
• Support for FaceNet architecture
• Added Embedding overlay plugin for face detection visualization

Known issues

• NCSDK does not support multiple calls to the inference engine from the same thread. This causes transitions when using the NCSDK backend to fail after the second play.
• Changing the backend on a stoped pipeline will fail with a segmentation fault.

v0.5.0

Introduced features:

• Support for the following architectures:
  • ResNet50V1
  • Tiny Yolo V3
  • FaceNetV1
  • MobileNetV2

• Support for the following backends:
  • NCSDK
  • TensorFlow

• Support for the following platforms:
• Intel Movidius Neural Compute Stick (version 1)
• NVIDIA Jetson AGX Xavier
• x86 systems
• NVIDIA TX2
• i.MX8

Known issues:

• NCSDK does not support multiple calls to the inference engine from the same thread. This causes the NCSDK backend to fail after the second start.
• Changing the backend on a stoped pipeline will fail with a segmentation fault.

v0.6.0

Introduced features:

• Improved ResNet50V1 inference result
• Preprocess and Postprocess factorized as general files.
• Debug information factorized as a general file.
• Tests added.
• Improved internal caps negotiation on bypass pad.
• Copyright license LGPL added.

Supported platforms:

• Intel Movidius Neural Compute Stick (version 1)
• NVIDIA Jetson AGX Xavier
• NVIDIA TX2
• NVIDIA Nano
• x86 systems
• i.MX8

Known issues:

• NCSDK does not support multiple calls to the inference engine from the same thread. This causes the NCSDK backend to fail after the second start.

v0.7.0

Introduced features:

• Pkg-config support
• License update to LGPL
• New inference meta hierarchical approach
• TFlite backend support
• New elements using new meta:
  • Detectioncrop
  • Inferencefilter
  • Inferencedebug
  • Inferenceoverlay
• Bug fixes

v0.7.1

Introduced features:

• Revert hotfix in dev-0.6 that modify the number of predictions.

v0.8.0

Introduced features:

• Add new inferenceutils plugin to absorb inferencecrop, inferencefilter and inferencedebug.
• Show root prediction on inferenceoverlay.
• Fix prediction size test.
• Fix tinyyolo3 postprocess to use new meta.
• Bug fixes

v0.9.0

Introduced features:

• Fix Yolo probabilities in inferencemeta
• Support for OpenCV 4
• Support for doubles in backend properties
• New inferencebin helper element