Overview

Image Segmentation on Qualcomm® Hexagon™ is a comprehensive container solution for running real-time segmentation models on the QCS6490 platform. Designed with full DSP acceleration, it brings plug-and-play deployment for models like YOLOv8-seg and DeepLabv3+ MobileNet, pre-optimized for edge scenarios.

This container offers:

Dual Image Segmentation Workflows:
- Ultralytics Export: Use YOLOv8-native tools to export to TFLite for rapid prototyping
- AI Hub Conversion: Import optimized DeepLabv3+ MobileNet models directly from Qualcomm’s Hugging Face repository
Integrated Runtime Stack:
- Pre-installed support for QNN, SNPE, and LiteRT
- Includes GStreamer, OpenCV, and Python 3.10 for full inference pipeline development
Hardware-Accelerated Inference:
- INT8 inference on Hexagon™ DSP 770
- FP32 fallback and GPU acceleration via Adreno™ 643 GPU
Multi-Model Format Compatibility:
- Runs .tflite, .dlc, and .so formats natively with supported runtimes
Preconfigured Scripts & Utilities:
- advantech-coe-model-export.sh and advantech-aihub-model-export.sh for model conversion
- wise-bench.sh for validating runtime and AI environment
Ready for Industrial Edge Use Cases:
- Built for robotics, medical imaging, automotive vision, industrial inspection, and smart agriculture
- Designed for use on Advantech AOM-2721 with QCS6490 SoC
Seamless ROS Support:
- Compatible with Qualcomm Robotics Reference Distro with ROS 1.3-ver.1.1 for plug-and-play robotic integration

Container Demo

Edge-Ready Use Cases

Domain	Key Applications
Fitness & Rehabilitation	Real-time posture feedback, remote physical therapy monitoring
Automotive & Robotics	Scene and object segmentation for autonomous navigation and robotic interaction
Healthcare & Medical Imaging	Tumor/organ segmentation and quantitative analysis for diagnostics
Satellite & Environmental Monitoring	Land cover classification, disaster impact detection, climate analysis
Smart Agriculture	Crop health monitoring, yield estimation, precision weed detection
Industrial Inspection	Automated defect detection on parts, PCBs, and production lines
Retail, eCommerce & AR	Virtual try-on, product isolation, visual search, background removal
Photography & AR	Portrait mode, background replacement, real-time visual effects
Bio-Imaging & Research	Cell and subcellular segmentation for life science research
Marine & Environmental Science	Coral reef monitoring, shoreline mapping, erosion analysis

Host Device Prerequisites

Component	Specification
Target Hardware	Advantech AOM-2721
SoC	Qualcomm® QCS6490
GPU	Adreno™ 643
DSP	Hexagon™ 770
Memory	8GB LPDDR5
Host OS	QCOM Robotics Reference Distro with ROS 1.3-ver.1.1
BSP	Yocto 4.0 (LE1.3)

Container Environment Overview

Software Components on Container Image

Component	Version	Description
LiteRT	1.3.0	Provides QNN TFLite Delegate support for GPU and DSP acceleration
SNPE	2.29.0	Qualcomm’s Snapdragon Neural Processing Engine; optimized runtime for Snapdragon DSP/HTP
QNN	2.29.0	Qualcomm® Neural Network (QNN) runtime for executing quantized neural networks
GStreamer	1.20.7	Multimedia framework for building flexible audio/video pipelines
Python	3.10.12	Python runtime for building applications
OpenCV	4.11.0	Computer vision library for image and video processing

Quick Start Guide

For container quick start, including the docker-compose file and more, please refer to README.

Supported AI Capabilities

Vision Models

Model	Format	Note
YOLOv8 Detection	TFLite INT8	Downloaded from Ultralytics` official source and exported to TFLite using Ultralytics Python packages
YOLOv8 Segmentation	TFLite INT8	Downloaded from Ultralytics` official source and exported to TFLite using Ultralytics Python packages
YOLOv8 Pose Estimation	TFLite INT8	Downloaded from Ultralytics` official source and exported to TFLite using Ultralytics Python packages
Lightweight Face Detector	TFLite INT8	Converted using Qualcomm® AI Hub
FaceMap 3D Morphable Model	TFLite INT8	Converted using Qualcomm® AI Hub
DeepLabV3+ (MobileNet)	TFLite INT8	Converted using Qualcomm® AI Hub
DeepLabV3 (ResNet50)	SNPE DLC TFLite	Converted using Qualcomm® AI Hub
HRNet Pose Estimation (INT8)	TFLite INT8	Converted using Qualcomm® AI Hub
PoseNet (MobileNet V1)	TFLite	Converted using Qualcomm® AI Hub
MiDaS Depth Estimation	TFLite INT8	Converted using Qualcomm® AI Hub
MobileNet V2 (Quantized)	TFLite INT8	Converted using Qualcomm® AI Hub
Inception V3 (SNPE DLC)	SNPE DLC TFLite	Converted using Qualcomm® AI Hub
YAMNet (Audio Classification)	TFLite	Converted using Qualcomm® AI Hub
YOLO (Quantized)	TFLite INT8	Converted using Qualcomm® AI Hub

Supported AI Model Formats

Runtime	Format	Compatible Versions
QNN	.so	2.29.0
SNPE	.dlc	2.29.0
LiteRT	.tflite	1.3.0

Hardware Acceleration Support

Accelerator	Support Level	Compatible Libraries
GPU	FP32	QNN, SNPE, LiteRT
DSP	INT8	QNN, SNPE, LiteRT

Best Practices

Prefer INT8 quantized models for DSP acceleration
Ensure fixed batch sizes when converting models
Use lower GST_DEBUG levels for stable multimedia handling
Always validate exported models on-device after deployment