Application development with QIM SDK - Qualcomm Dragonwing Documentation

Developing an AI application with QIM SDK consists of two steps: understanding the core components of the AI pipeline and following the application development workflow to integrate, configure, and deploy your application.

Important components

The four plugins every AI pipeline is built from, and which one usually blocks integration.

Development workflow

Use case → AI Hub model check → post-processing check → build the application.

Post-processing modules

Supported use cases, model families, and the metadata each module type produces.

Important QIM SDK components for an AI application

Every QIM SDK AI application is assembled from a small set of components. Two of them decide whether your use case works out of the box: the inference plugin that runs your model, and the post-processing module that turns raw output tensors into usable metadata.

Component	Plugin	What it does	Why it is important
AI preprocessing	`qtimlvconverter`	Resize, color conversion, and tensor-layout transformation on the GPU.	Produces the exact tensor your model expects, without making the CPU the bottleneck.
AI inference	`qtimltflite` · `qtimlqnn` · `qtimlsnpe`	Runs the model and returns raw output tensors.	Selects the Qualcomm acceleration path — LiteRT/TFLite delegate, Qualcomm AI Engine Direct/QNN, or SNPE. Your model format decides which plugin you use.
Post-processing	`qtimlpostprocess`	Loads a module that decodes output tensors into ML metadata — labels, boxes, masks, keypoints.	If no module matches your model’s output, the pipeline produces no usable results. This is the most common integration blocker.
Metadata consumption	`qtivoverlay` · application logic	Draws metadata on the frame with GPU acceleration, or hands it to your code.	Turns detections into what the product actually does: visualization, events, analytics, or control logic.

Post-processing is model-specific, not use-case-generic. Two object detectors can require different decode logic. Confirm a module exists for your exact model output before you plan the build — see Post-processing modules and supported use cases.

To list the QIM SDK plugins on your device, and inspect the post-processing modules and tensor shapes each one accepts:

gst-inspect-1.0 | grep qti

gst-inspect-1.0 qtimlpostprocess

Application development workflow

The workflow is driven by two decisions: whether a compatible model is available on Qualcomm AI Hub and whether an existing post-processing module supports the model output. If both are supported, proceed directly to application development. Otherwise, integrate your own model and/or add a custom post-processing module first.

Step	What you do	Where to go
1	Select the use case. Decide what the application must detect, classify, segment, or enhance, and note the required resolution and frame rate.	Supported use cases
2	Check if the model is available on Qualcomm AI Hub. If a suitable pre-optimized model exists, download it for your device and runtime.	Qualcomm AI Hub
2a	Bring your own model — only if no AI Hub model fits. Train or source the model, then compile and quantize it for the target.	Integrating a custom AI model
3	Check if post-processing is supported in QIM SDK. Compare your model’s output tensors with the available `qtimlpostprocess` modules.	Post-processing modules
3a	Add post-processing support — only if no module matches. Write, build, and deploy a custom module.	Add post-processing support
4	Build the application. Choose a development path, validate the pipeline, then build with GStreamer, C++, or Python App Builder.	Build the application

Steps 2a and 3a are optional. When a suitable AI Hub model and a compatible post-processing module are available, you can move directly from defining the use case to building the application.

Post-processing modules and supported use cases

The qtimlpostprocess plugin converts raw model output tensors into structured metadata that can be consumed by the rest of the pipeline. Configure the required post-processing module using the module property and provide any model-specific labels or settings needed by that module. Selecting a module that matches your model’s output is the single most important decision for whether the pipeline produces usable results.

Module type	Example model families	Raw model output	Metadata produced for downstream stages
`image-classification`	MobileNet and similar classifiers	Confidence scores per class	Top labels and scores, filtered by confidence threshold
`object-detection`	SSD-MobileNet, YOLOv5, YOLO-NAS, YOLOv8	Box coordinates, class IDs, scores	Decoded bounding boxes, mapped labels, and filtered confidence scores
`image-segmentation`	DeepLabV3, FFNet	Pixel-level class maps or depth tensors	Segmentation masks aligned with the source frame
`pose-estimation`	PoseNet MobileNet and similar keypoint models	Keypoint coordinates and confidences	Keypoints plus the links between them, ready to render or track
`super-resolution`	QuickSRNet, XLSR	Enhanced image tensors	Higher-resolution frames or tensors for the output stage
`audio-classification`	YAMNet and similar audio models	Audio class probabilities	Class labels and confidence scores
Raw tensor or custom output	Any model whose output no module matches	Model-specific tensor data	Nothing, until you provide a matching module — see Add post-processing support

The examples shown are representative model families rather than an exhaustive list. Since post-processing modules are generally based on output tensor formats, many compatible models can reuse the same module. Verify your model’s output tensors against the supported formats reported by gst-inspect-1.0 qtimlpostprocess before deciding that a custom module is required.

If your model falls into the raw tensor or custom output row, the pipeline will not produce usable metadata until a matching module exists. Plan for this before integration rather than during bring-up.

Next steps

Pick the card that matches where you are in the workflow.

Bring your own model

Step 2a — no AI Hub model fits, so compile and quantize your own for the target.

Add post-processing support

Step 3a — no built-in module matches your output, so write and deploy a custom module.

Build the application

Step 4 — model and module are ready, so choose a path and build.

Run the post-processing check (step 3) before you invest in module development. If an existing module already accepts your model’s output tensors, you skip step 3a entirely.

Qualcomm AI Hub — download pre-trained, optimized models to use in these pipelines.
Plugin reference — every QIM SDK plugin, with properties and caps.
Sample pipelines — ready-to-run pipelines for each use case.

Important components

Development workflow

Post-processing modules

​Important QIM SDK components for an AI application

​Application development workflow

​Post-processing modules and supported use cases

​Next steps

Bring your own model

Add post-processing support

Build the application

​Related information

Important QIM SDK components for an AI application

Application development workflow

Post-processing modules and supported use cases

Next steps

Related information