EdgeFirst AI

AI for Spatial Perception

EdgeFirst Perception is an open-source suite of libraries and microservices for AI-driven spatial perception on edge devices. It supports cameras, LiDAR, radar, and time-of-flight sensors — enabling real-time object detection, segmentation, sensor fusion, and 3D spatial understanding, optimized for resource-constrained embedded hardware.

Workflow

Every model in the EdgeFirst Model Zoo passes through a validated pipeline. EdgeFirst Studio manages datasets, training, multi-format export (ONNX, TFLite INT8, eIQ Neutron, Kinara DVM, HailoRT HEF, TensorRT), and reference validation. Models are then deployed to our board farm for full-dataset on-target validation on real hardware — measuring both accuracy (mAP) and detailed timing breakdown per device. Results are published here on HuggingFace with per-platform performance tables.

Model Lifecycle

On-Target Validation

Unlike desktop-only benchmarks, EdgeFirst validates every model on real target hardware with the full dataset. Each device produces both accuracy metrics (mAP) and a detailed timing breakdown — load, preprocessing, NPU inference, and decode — so you know exactly how a model performs on your specific platform.

Supported Hardware

Model Zoo

Pre-trained YOLO models for edge deployment. Each model repo contains all sizes (nano through x-large), ONNX FP32 and TFLite INT8 formats, with platform-specific compiled variants as they become available.

Detection

YOLO26

n/s/m/l/x · COCO 80 classes · Nano mAP@0.5: 54.9%

YOLO11

n/s/m/l/x · COCO 80 classes · Nano mAP@0.5: 53.4%

YOLOv8

n/s/m/l/x · COCO 80 classes · Nano mAP@0.5: 50.2%

YOLOv5

n/s/m/l/x · COCO 80 classes · Nano mAP@0.5: 49.6%

Instance Segmentation

YOLO26

n/s/m/l/x · COCO 80 classes · Nano Mask mAP: 37.0%

YOLO11

n/s/m/l/x · COCO 80 classes · Nano Mask mAP: 35.5%

YOLOv8

n/s/m/l/x · COCO 80 classes · Nano Mask mAP: 34.1%

Roadmap

The EdgeFirst Model Zoo is expanding across the full spatial perception stack — from 2D detection through depth estimation, 3D scene understanding, and edge VLMs. All models are validated on real hardware with the same pipeline used for our YOLO models.

Category	Examples	Platforms	Status
Detection (Apache 2.0)	DETR-class, EfficientDet, mobile-optimized detectors	i.MX Ara240 Hailo Jetson	Coming Soon
Semantic Segmentation	Lightweight real-time scene parsing	i.MX Ara240 Hailo Jetson	Roadmap
Instance Segmentation (Apache 2.0)	Non-YOLO mask prediction	Ara240 Jetson	Roadmap
SAM-like Segmentation	Prompted, class-agnostic masks	Ara240 Jetson	Roadmap
Monocular Depth	Relative and metric depth estimation	i.MX Ara240 Jetson	Roadmap
Stereo Depth	Hardware stereo depth matching	Jetson	Roadmap
3D Detection & Occupancy	Monocular 3D, BEV, occupancy grids	Jetson	Roadmap
Edge VLMs	Visual language models for edge inference	Ara240 Jetson	Roadmap

Roadmap is subject to change. Models are published as validation completes on each target platform.

Naming Convention

Each HuggingFace repo contains one model family for one task, with all size variants inside.

Component	Pattern	Example
HF Repo	EdgeFirst/{version}-{task}	EdgeFirst/yolov8-det
ONNX Model	{version}{size}-{task}.onnx	yolov8n-det.onnx
TFLite Model	{version}{size}-{task}-int8.tflite	yolov8n-det-int8.tflite
i.MX 95 TFLite	{version}{size}-{task}.imx95.tflite	yolov8n-det.imx95.tflite
i.MX 93 TFLite	{version}{size}-{task}.imx93.tflite	yolov8n-det.imx93.tflite
i.MX 943 TFLite	{version}{size}-{task}.imx943.tflite	yolov8n-det.imx943.tflite
Hailo HEF	{version}{size}-{task}.hailo{variant}.hef	yolov8n-det.hailo8l.hef
Studio Project	{Dataset} {Task}	COCO Detection
Studio Experiment	{Version} {Task}	YOLOv8 Detection

Validation Pipeline

Models go through two validation stages before publication:

Stage	What	Where
Reference	ONNX FP32 and TFLite INT8 mAP on full COCO val2017 (5000 images)	EdgeFirst Studio (cloud)
On-Target	Full dataset mAP + timing breakdown (load, preproc, invoke, decode, e2e) per device	Board farm (real hardware) In Progress

Perception Architecture

Layer	Description
Foundation	Hardware abstraction, video I/O, accelerated inference delegates
Zenoh	Modular perception pipeline over Zenoh pub/sub
GStreamer	Spatial perception elements for GStreamer / NNStreamer
ROS 2	Native ROS 2 nodes extending Zenoh microservices Roadmap

EdgeFirst Studio

EdgeFirst Studio is the MLOps platform that drives the entire model zoo pipeline. Free tier available.

Dataset management & AI-assisted annotation
Model training with automatic multi-format export and INT8 quantization
Reference and on-target validation with full metrics collection
CameraAdaptor integration for native sensor format training
Deploy trained models to edge devices via the EdgeFirst Client CLI