Download source code for development - Qualcomm Dragonwing Documentation

The Standard SDK (SDK) is required to develop application and plugin code. See the Concepts of Building Native GStreamer Application for instructions on setting up the SDK and downloading and compiling the source code.

Qualcomm IM SDK preprocessing and postprocessing ML plugins

AI/ML applications are built from the IM SDK’s preprocessing (qtimlvconverter), inference (qtimltflite, qtimlqnn, qtimlonnx, qtimlsnpe), and post-processing (qtimlpostprocess) plugins. For how these map to the pipeline and hardware engines, see the IM SDK overview; for the complete plugin catalog, see the plugin reference in Discover SDKs. Preprocessing plugin

Plugin	Functionality
`qtimlvconverter`	Transforms incoming video buffers into neural-network tensors while performing required format conversion and resizing.

Postprocessing plugin

Plugin	Functionality
`qtimlpostprocess`	A customizable plugin that provides a library interface for postprocessing the tensor output of inference plugins.

The following module types are supported by the qtimlpostprocess plugin. Each module type handles postprocessing for a specific use case.

Module Type	Functionality
`audio-classification`	Performs postprocessing of output tensors for audio classification use cases.
`image-classification`	Performs postprocessing of output tensors for image classification use cases.
`image-segmentation`	Performs postprocessing of output tensors for pixel-level use cases such as image segmentation and depth mapping.
`object-detection`	Performs postprocessing of output tensors for object detection use cases.
`pose-estimation`	Performs postprocessing of output tensors for pose estimation use cases.
`super-resolution`	Performs postprocessing of output tensors for video super resolution use cases.

The qtimlpostprocess plugin supports the following use cases and related models.

Use cases supported by Qualcomm IM SDK	Supported Models
Classification	Models like Mobilenet. Currently Qualcomm AI Hub has 11 classification models supported. New models will keep getting added to AI Hub.
Detection	Models like ssd-mobilenet, yolov5, yolo-nas, and yolov8
Segmentation	Models like deeplabv3_resnet and ffnet
Pose detection	Models like posenet_mobilenet
Super resolution	Models like QuickSRNet, XLSR, etc.

For verified Qualcomm AI Hub models and ready-to-run pipeline examples, see Discover SDKs → IM SDKs.You can use many other models with similar postprocessing requirements. However, it is recommended to verify postprocessing support in the relevant Qualcomm IM SDK plugins before integrating your own model.