QIM SDK Development Guide

The Qualcomm Intelligent Multimedia SDK (IM SDK) is a GStreamer-based AI/ML application development framework. It provides ready-to-use CLI tools and hardware-accelerated plugins covering the complete pipeline from video capture, AI inference, to result rendering.

Relationship with QAIRT

QIM SDK sits on top of QAIRT SDK:

Layer	Contents	Status on Q900
QAIRT Runtime	QNN / SNPE / LiteRT runtime libraries	Installed via apt (qairt-libs)
QIM SDK Plugins	GStreamer ML plugins (capture, preprocess, inference, postprocess, render)	Installed via apt
QIM SDK Sample Apps	gst-ai-* CLI tools	Installed via apt

Install QAIRT base libraries first (sudo apt install -y qairt-libs), see the installation steps below.

Prerequisites

• Radxa AIRbox Q900 (QCS9075 / SA8775P, HTP V73 NPU)
• Ubuntu 24.04, with network access
• Display connected (Wayland)

Installation

Step 1: Install QAIRT Base Libraries

radxa@airbox$

sudo apt install -y qairt-libs

Step 2: Install QIM SDK

radxa@airbox$

sudo apt install -y \
    gstreamer1.0-tools \
    tensorflow-lite-qcom-apps \
    gstreamer1.0-qcom-sample-apps \
    gstreamer1.0-qcom-python-examples

All ML inference plugins and video processing plugins are automatically installed as dependencies.

Step 3: Verify Installation

radxa@airbox$

ls /usr/bin/gst-ai-*

Expected output: 20 gst-ai-* CLI tools.

Download Models and Test Resources

Use the official script for one-click download:

radxa@airbox$

sudo apt install -y unzip
curl -L -O https://raw.githubusercontent.com/quic/sample-apps-for-qualcomm-linux/refs/heads/main/download_artifacts.sh
chmod +x download_artifacts.sh
sudo ./download_artifacts.sh -v GA1.6-rel -c QCS9075

After download:

Directory	Contents
/etc/models/	10 .tflite quantized models
/etc/labels/	16 label and config files
/etc/media/	4 test videos
/etc/configs/	38 app config JSONs (from package install)

Quick Verification

radxa@airbox$

gst-ai-classification --config-file=/etc/configs/config_classification.json

On success, the display shows the test video with classification labels overlaid. The terminal output includes Using DSP Delegate and Pipeline state changed to PLAYING.

Press Ctrl + C to stop.

Normally you do not need to manually set XDG_RUNTIME_DIR or WAYLAND_DISPLAY — these are configured automatically on SSH login.

Available Sample Applications

The following apps have been verified on Q900 (QCS9075):

Application	Description	Model	NPU
Image Classification	Identify subjects in images	InceptionV3	DSP
Object Detection	Detect and locate objects	YOLOX	DSP
Pose Detection	Human body keypoint detection	HRNet	DSP
Image Segmentation	Per-pixel semantic segmentation	DeepLabV3+	DSP
Monocular Depth	Depth heatmap estimation	MiDaS V2	DSP
Super Resolution	Low-resolution upscaling	QuickSRNet	DSP (partial)
Audio Classification	Audio event classification	YAMNet	CPU
Face Detection	Face landmark detection	Lightweight Face Detection	DSP
Daisy-Chain Detection & Classification	Cascaded detection + classification	YOLOX + InceptionV3	DSP
Daisy-Chain Detection & Pose	Cascaded detection + pose	YOLOX + HRNet	DSP
Parallel AI Fusion	4-model simultaneous inference	YOLOX + InceptionV3 + HRNet + DeepLabV3+	DSP
Multistream Inference	Multi-stream concurrent detection	YOLOX	DSP
Event Encoder	Detection event encoding	YOLOX	DSP
Metadata Parser	Detection metadata export	YOLOX	DSP

See the IM SDK Reference Manual Chapter 3 for the full list.

Reference

• Python GStreamer Apps — Python binding examples
• Build from Source — Compile custom GStreamer apps on-device

Troubleshooting

No display output

Check that the Wayland socket exists:

radxa@airbox$

ls $XDG_RUNTIME_DIR/$WAYLAND_DISPLAY

DSP inference failure

Verify FastRPC device nodes exist:

radxa@airbox$

ls /dev/fastrpc-cdsp*

If no output, see Enable NPU.

Model files missing

Re-run the download script:

radxa@airbox$

sudo ./download_artifacts.sh -v GA1.6-rel -c QCS9075