Edge Computing

Real-Time AI, In Your Browser

YOLO Edge Runner brings YOLO11 object detection, segmentation, and pose estimation to the browser. WebGPU-accelerated inference with zero server overhead — your data never leaves the device.

Open App Source

~290msInference (GPU)

0msServer Latency

60fpsUI Refresh

Supported Tasks

Three computer vision tasks, all running entirely in-browser with a single unified pipeline.

Object Detection

Identify and locate objects in images with bounding boxes and confidence scores. Supports 80 COCO classes out of the box.

Instance Segmentation

Pixel-level masks for each detected object. Separate overlapping instances with distinct color-coded regions.

Pose Estimation

17-keypoint skeleton detection per person. Tracks body joints including eyes, shoulders, elbows, wrists, hips, knees, and ankles.

System Architecture

Non-Blocking Worker Pattern

To maintain a smooth 60 FPS UI, inference runs off the main thread. A dedicated Web Worker handles the entire ONNX lifecycle independently.

Execution Provider (EP)

The engine benchmarks your hardware to select the best provider: WebGPU for modern GPUs, or multi-threaded WASM for cross-compatibility.

Memory Management

Transferable Objects move pixel data between threads with zero-copy overhead, ensuring maximum throughput for high-resolution cameras.

Technology

Built on the bleeding edge of the Open Web platform.

YOLO11

Ultralytics Engine

ONNX Runtime

Microsoft AI Core

Next.js 16

App Router / RSC

WebGPU

W3C Next-Gen Graphics

FP16 & INT8 Quantization

To run efficiently at the edge, we use Half-Precision (FP16) calibration. This reduces model size by 50% while leveraging hardware-level bit-shifting on modern GPUs via the shader-f16 extension.

50% Memory ReductionWebGPU Throughput Acceleration

Precision Utility

/** Bit-depth Transformation Utility */
function encodeFloat16(val) {
  // IEEE 754 float32 → float16
  exponent = exponent - 127 + 15;
  return sign | (exponent << 10)
    | (mantissa >> 13);
}

Attribution

Built with open-source tools from incredible teams.

Ultralytics

Engine

YOLO11 is an object detection, segmentation, and pose estimation model. Thanks to the Ultralytics team for their open-source contribution to the computer vision community.

ultralytics.com

ONNX Runtime

Microsoft

Microsoft's ONNX Runtime provides the high-performance WebGPU kernels that power inference in this application.

onnxruntime.ai

Multi-Task Web

Reference

Huge credits to nomi30701 for their fantastic repository which served as a crucial foundation and reference for the multi-task web implementation.

GitHub

Developed by Pranta Barua