Hugging Face's logo

Spaces:
EdgeFirst
/
Models
Running

App Files Community
Models / index.html
sebastientaylor's picture
sebastientaylor
Upload index.html with huggingface_hub
53a527f verified
Raw
History Blame
47.6 kB
<!DOCTYPE html>
<html lang="en">
<head>
 <meta charset="UTF-8">
 <meta name="viewport" content="width=device-width, initial-scale=1.0">
 <title>EdgeFirst AI — Model Zoo</title>
 <base target="_blank">
 <link rel="preconnect" href="https://fonts.googleapis.com">
 <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
 <link href="https://fonts.googleapis.com/css2?family=Barlow:wght@300;400;500;600;700&family=Crimson+Text:wght@400;600&family=JetBrains+Mono:wght@400;500&display=swap" rel="stylesheet">
 <style>
 :root {
 --navy: #3E3371;
 --gold: #E8B820;
 --teal: #1FA0A8;
 --teal-text: #167A80;
 --indigo: #4B0082;
 --blue: #8FA3D4;
 --bg: #FFFFFF;
 --bg-subtle: #F8F9FA;
 --bg-card: #F0EDF8;
 --text: #343A40;
 --text-strong: #212529;
 --text-muted: #6C757D;
 --border: #E9ECEF;
 --heading: var(--navy);
 --link: var(--teal-text);
 }
 @media (prefers-color-scheme: dark) {
 :root {
 --bg: #1a1a2e;
 --bg-subtle: #16213e;
 --bg-card: rgba(75, 0, 130, 0.2);
 --text: #F1F3F5;
 --text-strong: #FFFFFF;
 --text-muted: #aaa;
 --border: rgba(255,255,255,0.1);
 --heading: #FFFFFF;
 --link: #1FA0A8;
 }
 }
 * { margin: 0; padding: 0; box-sizing: border-box; }
 html { scroll-behavior: smooth; }
 body {
 font-family: 'Crimson Text', Georgia, serif;
 background: var(--bg);
 color: var(--text);
 line-height: 1.7;
 font-size: 16px;
 }
 .container { max-width: 900px; margin: 0 auto; padding: 2.5rem 2rem; }
 h1, h2, h3, h4 {
 font-family: 'Barlow', -apple-system, sans-serif;
 color: var(--heading);
 line-height: 1.3;
 }
 h1 { font-size: 2.2rem; font-weight: 700; margin-bottom: 0.15rem; }
 h1 .edge { color: var(--heading); }
 h1 .first { color: var(--gold); }
 h2 {
 font-size: 1.15rem; font-weight: 600;
 text-transform: uppercase; letter-spacing: 0.06em;
 margin-top: 2.5rem; margin-bottom: 0.75rem;
 padding-bottom: 0.35rem; border-bottom: 2px solid var(--gold);
 }
 h3 {
 font-size: 1.05rem; font-weight: 600;
 margin-top: 1.75rem; margin-bottom: 0.5rem;
 }
 h4 {
 font-size: 0.9rem; font-weight: 600;
 color: var(--text-muted);
 text-transform: uppercase; letter-spacing: 0.04em;
 margin-top: 1.25rem; margin-bottom: 0.4rem;
 }
 .tagline {
 font-family: 'Barlow', sans-serif; font-weight: 500;
 letter-spacing: 0.12em; text-transform: uppercase;
 color: var(--text-muted); font-size: 0.85rem; margin-bottom: 1.5rem;
 }
 p { margin-bottom: 1rem; }
 a { color: var(--link); text-decoration: none; }
 a:hover { text-decoration: underline; color: var(--navy); }
 @media (prefers-color-scheme: dark) { a:hover { color: var(--gold); } }
 /* Navigation */
 .toc {
 font-family: 'Barlow', sans-serif;
 background: var(--bg-subtle);
 border: 1px solid var(--border);
 border-radius: 6px;
 padding: 1rem 1.25rem;
 margin: 1.5rem 0;
 font-size: 0.88rem;
 }
 .toc-title {
 font-weight: 600; font-size: 0.78rem;
 text-transform: uppercase; letter-spacing: 0.05em;
 color: var(--text-muted); margin-bottom: 0.5rem;
 }
 .toc ul { list-style: none; display: flex; flex-wrap: wrap; gap: 0.4rem 1.5rem; }
 .toc a { font-weight: 500; }
 .breadcrumb {
 font-family: 'Barlow', sans-serif;
 font-size: 0.78rem; color: var(--text-muted);
 margin-bottom: 0.5rem;
 }
 .breadcrumb a { color: var(--text-muted); font-weight: 500; }
 .back-to-top {
 font-family: 'Barlow', sans-serif;
 font-size: 0.75rem; color: var(--text-muted);
 display: inline-block; margin-top: 0.5rem;
 }
 /* Badges */
 .badges { display: flex; flex-wrap: wrap; gap: 0.4rem; margin: 1.25rem 0; }
 .badges img { height: 22px; }
 .link-badges { display: flex; flex-wrap: wrap; gap: 0.4rem; margin: 1.5rem 0; }
 .link-badges img { height: 28px; }
 .platform-badge {
 font-family: 'Barlow', sans-serif;
 font-size: 0.62rem; font-weight: 600;
 padding: 0.06rem 0.3rem; border-radius: 2px;
 letter-spacing: 0.02em; white-space: nowrap;
 background: var(--bg-card); color: var(--text-muted);
 border: 1px solid var(--border);
 }
 .status-available {
 font-family: 'Barlow', sans-serif;
 font-size: 0.68rem; font-weight: 600;
 padding: 0.1rem 0.4rem; border-radius: 3px;
 background: #1B7F37; color: #fff;
 letter-spacing: 0.03em; text-transform: uppercase;
 }
 .status-coming {
 font-family: 'Barlow', sans-serif;
 font-size: 0.68rem; font-weight: 600;
 padding: 0.1rem 0.4rem; border-radius: 3px;
 background: var(--gold); color: #333;
 letter-spacing: 0.03em; text-transform: uppercase;
 }
 .status-planned {
 font-family: 'Barlow', sans-serif;
 font-size: 0.68rem; font-weight: 600;
 padding: 0.1rem 0.4rem; border-radius: 3px;
 background: var(--indigo); color: #fff;
 letter-spacing: 0.03em; text-transform: uppercase;
 }
 .roadmap-tag {
 font-family: 'Barlow', sans-serif;
 font-size: 0.7rem; font-weight: 600;
 padding: 0.1rem 0.45rem; border-radius: 3px;
 background: var(--indigo); color: #fff;
 vertical-align: middle; margin-left: 0.3rem;
 letter-spacing: 0.03em; text-transform: uppercase;
 }
 .wip-tag {
 font-family: 'Barlow', sans-serif;
 font-size: 0.7rem; font-weight: 600;
 padding: 0.1rem 0.45rem; border-radius: 3px;
 background: var(--gold); color: #333;
 vertical-align: middle; margin-left: 0.3rem;
 letter-spacing: 0.03em; text-transform: uppercase;
 }
 /* Diagrams */
 .diagram-container { margin: 1.5rem 0; text-align: center; }
 .diagram-container img { max-width: 100%; height: auto; }
 /* Tables */
 .data-table {
 width: 100%;
 border-collapse: collapse;
 margin: 0.75rem 0;
 font-family: 'Barlow', sans-serif;
 font-size: 0.88rem;
 }
 .data-table th {
 text-align: left;
 padding: 0.5rem 0.6rem;
 font-weight: 600; font-size: 0.75rem;
 text-transform: uppercase; letter-spacing: 0.04em;
 color: var(--text-muted);
 border-bottom: 2px solid var(--border);
 }
 .data-table td {
 padding: 0.4rem 0.6rem;
 border-bottom: 1px solid var(--border);
 }
 .data-table td:first-child {
 font-weight: 600;
 color: var(--text-strong);
 white-space: nowrap;
 }
 .data-table code {
 font-family: 'JetBrains Mono', monospace;
 font-size: 0.78rem;
 }
 .data-table .num { font-family: 'JetBrains Mono', monospace; font-size: 0.82rem; }
 .data-table .muted { color: var(--text-muted); }
 .footnote {
 color: var(--text-muted); font-size: 0.82rem;
 margin: 0.5rem 0 1.5rem 0; line-height: 1.5;
 }
 .footnote a { color: var(--text-muted); }
 .drill-down {
 background: var(--bg-card);
 border-left: 4px solid var(--gold);
 border-radius: 6px;
 padding: 1rem 1.2rem;
 margin: 1.2rem 0;
 }
 .drill-down p { margin: 0.3rem 0; font-size: 0.92rem; }
 .drill-down .cta-link {
 display: inline-block;
 margin-top: 0.4rem;
 font-weight: 600;
 color: var(--navy);
 }
 @media (prefers-color-scheme: dark) {
 .drill-down .cta-link { color: var(--gold); }
 }
 /* Model cards */
 .model-grid {
 display: grid;
 grid-template-columns: repeat(auto-fill, minmax(240px, 1fr));
 gap: 0.75rem; margin: 1rem 0;
 }
 .model-card {
 font-family: 'Barlow', sans-serif;
 background: var(--bg-card);
 padding: 0.85rem 1rem; border-radius: 5px;
 border-left: 3px solid var(--gold);
 transition: border-color 0.15s;
 }
 .model-card:hover { border-left-color: var(--teal); }
 .model-card h3 { font-size: 0.95rem; font-weight: 600; margin: 0 0 0.2rem 0; }
 .model-card h3 a { color: var(--text-strong); }
 .model-card h3 a:hover { color: var(--link); text-decoration: none; }
 .model-card .meta { color: var(--text-muted); font-size: 0.82rem; font-weight: 400; }
 /* Accuracy bars */
 .bar-chart { margin: 0.75rem 0; }
 .bar-row {
 display: flex; align-items: center; gap: 0.5rem;
 font-family: 'Barlow', sans-serif; font-size: 0.82rem;
 margin-bottom: 0.3rem;
 }
 .bar-label { width: 65px; font-weight: 600; color: var(--text-strong); text-align: right; }
 .bar-track { flex: 1; height: 18px; background: var(--bg-subtle); border-radius: 3px; overflow: hidden; }
 .bar-fill { height: 100%; border-radius: 3px; display: flex; align-items: center; padding-left: 6px; }
 .bar-fill span { font-size: 0.72rem; font-weight: 600; color: #fff; white-space: nowrap; }
 .bar-navy { background: var(--navy); }
 .bar-teal { background: var(--teal); }
 .bar-gold { background: var(--gold); }
 .bar-gold span { color: #333; }
 /* Collapsible sections */
 details { margin: 0.5rem 0; }
 details summary {
 font-family: 'Barlow', sans-serif;
 font-weight: 600; font-size: 0.9rem;
 cursor: pointer; padding: 0.4rem 0;
 color: var(--text-strong);
 }
 details summary:hover { color: var(--link); }
 /* Roadmap */
 .roadmap-note {
 font-family: 'Barlow', sans-serif;
 font-size: 0.82rem; color: var(--text-muted);
 font-style: italic; margin-top: 0.75rem;
 }
 .badge-row { display: flex; flex-wrap: wrap; gap: 0.25rem; }
 /* Studio features */
 .studio-features { margin: 0.75rem 0 0 1.25rem; color: var(--text); font-size: 0.95rem; }
 .studio-features li { margin-bottom: 0.25rem; }
 /* Footer */
 .footer {
 margin-top: 3rem; padding-top: 1.5rem;
 border-top: 1px solid var(--border);
 text-align: center;
 font-family: 'Barlow', sans-serif;
 font-size: 0.8rem; color: var(--text-muted);
 }
 .footer a { color: var(--text-muted); }
 .footer a:hover { color: var(--link); }
 /* Family section divider */
 .family-section {
 margin-top: 2rem;
 padding-top: 1.5rem;
 border-top: 1px solid var(--border);
 }
 .family-section:first-of-type { border-top: none; padding-top: 0; }
 </style>
</head>
<body>
 <div class="container">
<!-- ═══════════════════════════════════════════════════════ -->
 <!-- LEVEL 1: OVERVIEW -->
 <!-- ═══════════════════════════════════════════════════════ -->
<h1 id="top"><span class="edge">Edge</span><span class="first">First</span> AI</h1>
 <p class="tagline">Model Zoo &mdash; Edge AI Perception Models</p>
<p>
 Pre-trained models optimized for edge deployment, validated on real hardware with full-dataset accuracy metrics and per-platform timing breakdowns. Each model repo contains all sizes (nano through x-large) with ONNX FP32, TFLite INT8, and platform-specific compiled formats.
 </p>
<div class="link-badges">
 <a href="https://edgefirst.studio"><img src="https://img.shields.io/badge/EdgeFirst_Studio-3E3371?style=for-the-badge&logoColor=white" alt="EdgeFirst Studio"></a>
 <a href="https://github.com/EdgeFirstAI"><img src="https://img.shields.io/badge/GitHub-212529?style=for-the-badge&logo=github&logoColor=white" alt="GitHub"></a>
 <a href="https://doc.edgefirst.ai"><img src="https://img.shields.io/badge/Documentation-1FA0A8?style=for-the-badge&logo=readthedocs&logoColor=white" alt="Documentation"></a>
 <a href="https://www.au-zone.com"><img src="https://img.shields.io/badge/Au--Zone_Technologies-6C757D?style=for-the-badge" alt="Au-Zone Technologies"></a>
 </div>
<nav class="toc">
 <div class="toc-title">Model Families</div>
 <ul>
 <li><a href="#yolo26" target="_self">YOLO26</a></li>
 <li><a href="#yolo11" target="_self">YOLO11</a></li>
 <li><a href="#yolov8" target="_self">YOLOv8</a></li>
 <li><a href="#yolov5" target="_self">YOLOv5</a></li>
 </ul>
 </nav>
<!-- ═══════════════════════════════════════════════════════ -->
 <!-- EXECUTIVE SUMMARY: VALIDATION PIPELINE + PLATFORM -->
 <!-- ═══════════════════════════════════════════════════════ -->
<h2>Validation Pipeline</h2>
 <div class="diagram-container">
 <img src="01-ecosystem.png" alt="EdgeFirst Model Zoo Ecosystem">
 </div>
 <p>
 Every artifact in the Model Zoo is measured on the <strong>same dataset on the same hardware</strong> users deploy on. <a href="https://edgefirst.studio"><strong>EdgeFirst Studio</strong></a> manages datasets, training, multi-format export, and reference validation; on-target runs happen on a board farm of i.MX 8M Plus, i.MX 95, Ara-240, Hailo, and Jetson devices. Accuracy numbers and per-stage timing are pushed back to Studio session metrics and consumed by this Model Zoo when generating each model card.
 </p>
<h3>End-to-end Flow</h3>
 <div class="diagram-container">
 <img src="02-model-lifecycle.png" alt="Model Lifecycle: 5 stages from training to publication">
 </div>
 <p>
 Each training session produces a single set of weights. The export pipeline emits ONNX FP32, INT8 TFLite, and platform-specific compiled formats (i.MX 95 Neutron, NXP Ara-240 DVM, Hailo HEF, Jetson TensorRT). Every output is paired with an on-target validation run that captures both <strong>accuracy</strong> (COCO/LVIS mAP against the validation set) and <strong>full-pipeline timing</strong>. The ONNX FP32 run from each training session serves as the reference baseline; quantization and runtime loss are measured relative to it, not relative to externally-published numbers.
 </p>
<h3>EdgeFirst Profiler</h3>
 <div class="diagram-container">
 <img src="03-on-target-validation.png" alt="On-Target Validation Pipeline">
 </div>
 <p>
 The <strong>EdgeFirst Profiler</strong> is the on-target agent that drives every validation. Given a model and a dataset, it runs full inference on the target device, captures per-image predictions in the EdgeFirst Arrow/Parquet format, and emits a <a href="https://perfetto.dev"><strong>Perfetto</strong></a> trace alongside the predictions. The Profiler is hardware-aware: it loads each runtime through its native delegate &mdash; Verisilicon VX Delegate on i.MX 8M Plus, eIQ Neutron Delegate on i.MX 95, Kinara SDK on Ara-240, HailoRT on RPi5 + Hailo, TensorRT on Jetson &mdash; so every stage timed by the trace corresponds to what a deployed application would experience.
 </p>
<h3>EdgeFirst Validator</h3>
 <p>
 The <strong>EdgeFirst Validator</strong> is the off-target post-processor. It consumes the Profiler's predictions and Perfetto trace, computes the full 12-metric COCO accuracy tuple via <code>pycocotools</code> (or <code>lvis-api</code> for large-vocabulary datasets), and rebuilds per-stage timing summaries from the trace. Results land in a structured YAML payload attached to the Studio validation session &mdash; the same payload the Model Zoo reads to render this page. Accuracy and timing are computed independently of the runtime that produced the predictions, so toolchain regressions surface as cross-platform divergence rather than silent failures.
 </p>
<h3>EdgeFirst HAL</h3>
 <p>
 The <a href="https://github.com/EdgeFirstAI/hal"><strong>EdgeFirst Hardware Abstraction Layer</strong> (HAL)</a> provides the hardware-accelerated primitives used at both validation and deployment time. The Profiler uses HAL for letterbox resize, color-space conversion, normalization, layout conversion, and post-decode (YOLO/ModelPack output decoding, NMS, mask materialisation). HAL automatically selects DMA-BUF, OpenGL ES, NXP G2D, or CPU paths depending on the platform &mdash; so the timing measured during validation reflects the same accelerated path a production runtime would take. HAL ships as a Rust crate, a Python package, and a C library under Apache 2.0.
 </p>
<h3>Latency and Pipelined Throughput</h3>
 <p>
 Two complementary timing surfaces are reported per validation:
 </p>
 <table class="data-table">
 <tr><th>Surface</th><th>What it captures</th><th>When it&#39;s present</th></tr>
 <tr><td><code>timing.inline</code></td><td>Per-image <code>preprocess_ms</code>, <code>inference_ms</code>, <code>postprocess_ms</code> with min / mean / median / p95 / p99 / max</td><td>Always &mdash; the universal contract every producer fills in</td></tr>
 <tr><td><code>timing.trace</code></td><td>Full per-stage breakdown from the Perfetto trace (typically 25&ndash;33 stages including delegate work, tensor moves, decode passes, NMS, parquet flush), plus end-to-end FPS distribution</td><td>When the Profiler emits a sidecar trace (almost all runs)</td></tr>
 </table>
 <p>
 <strong>Throughput exceeds the sum of stage latencies</strong> because the runtime pipelines I/O, host preprocessing, NPU inference, and decode across frames. Reporting <code>1000 / (preprocess + inference + postprocess)</code> understates real throughput; the Model Zoo uses the <strong>measured end-to-end FPS</strong> from the trace (<code>trace.fps.median</code>) as the headline number reported in every per-target table on every model card. As a concrete example, YOLOv5n on i.MX 95 Neutron has per-stage means 21.7 ms preprocess + 12.2 ms inference + 15.8 ms postprocess (naive estimate ~20 FPS), but the measured pipelined throughput is <strong>56 FPS median</strong> &mdash; the 2.8&times; gap is the value the pipelining delivers.
 </p>
<h2>Platform Support</h2>
 <div class="badges">
 <img src="https://img.shields.io/badge/NXP-i.MX_8M_Plus-3E3371?style=flat-square&logoColor=white" alt="NXP i.MX 8M Plus">
 <img src="https://img.shields.io/badge/NXP-i.MX_95-3E3371?style=flat-square&logoColor=white" alt="NXP i.MX 95">
 <img src="https://img.shields.io/badge/NXP-Ara240-3E3371?style=flat-square&logoColor=white" alt="NXP Ara240">
 <img src="https://img.shields.io/badge/RPi5-Hailo--8%2F8L-1FA0A8?style=flat-square&logoColor=white" alt="RPi5 + Hailo-8/8L">
 <img src="https://img.shields.io/badge/NVIDIA-Jetson-76B900?style=flat-square&logoColor=white" alt="NVIDIA Jetson">
 </div>
<table class="data-table">
 <tr><th>Family</th><th>ONNX</th><th>i.MX 8M Plus</th><th>i.MX 95</th><th>Ara240</th><th>Hailo</th><th>Jetson</th></tr>
 <tr><td>YOLO26</td><td>✓</td><td>✓ <sup><a href="#platform-notes" target="_self">[1]</a></sup></td><td>✓ <sup><a href="#platform-notes" target="_self">[1]</a></sup></td><td>✓</td><td>✓</td><td>✓</td></tr>
 <tr><td>YOLO11</td><td>✓</td><td>✓</td><td>✓ <sup><a href="#platform-notes" target="_self">[1]</a></sup></td><td>✓</td><td>✓</td><td>✓</td></tr>
 <tr><td>YOLOv8</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr>
 <tr><td>YOLOv5</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr>
 </table>
<p id="platform-notes" class="footnote">
 ✓ indicates a compiled model artifact is published in the corresponding repo. <sup>[1]</sup> Some (size, platform) combinations are work in progress and are not yet showing a number in the per-repo model card's <em>On-target validation results</em> table. Two reasons cover most cases: <em>timing</em> &mdash; larger sizes on slower NPUs are not yet a validation priority and will roll in as bandwidth allows; and <em>accuracy or performance investigations</em> &mdash; for example, YOLO11 and YOLO26 on the i.MX 95 Neutron currently show a quantization regression that we are tracking with NXP. In every case the underlying Studio validation session (<code>v-XXXX</code>) remains linked in the model card so its current status can be inspected.
 </p>
<!-- Overview: All families comparison -->
 <h2>Detection &mdash; Nano Accuracy</h2>
 <p>ONNX FP32 mAP@0.5 on COCO val2017 (5000 images, 80 classes). Nano size for each family. Source: EdgeFirst Studio validation sessions cited in each model card.</p>
 <div class="bar-chart">
 <div class="bar-row"><span class="bar-label">YOLO26</span><div class="bar-track"><div class="bar-fill bar-navy" style="width:91.8%"><span>55.06%</span></div></div></div>
 <div class="bar-row"><span class="bar-label">YOLO11</span><div class="bar-track"><div class="bar-fill bar-teal" style="width:88.4%"><span>53.05%</span></div></div></div>
 <div class="bar-row"><span class="bar-label">YOLOv8</span><div class="bar-track"><div class="bar-fill bar-navy" style="width:84.3%"><span>50.55%</span></div></div></div>
 <div class="bar-row"><span class="bar-label">YOLOv5</span><div class="bar-track"><div class="bar-fill bar-teal" style="width:79.6%"><span>47.77%</span></div></div></div>
 </div>
<h2>Segmentation &mdash; Nano Mask Accuracy</h2>
 <p>ONNX FP32 Mask mAP@0.5-0.95 on COCO val2017. Nano size, split-decoder. Source: EdgeFirst Studio validation sessions cited in each model card.</p>
 <div class="bar-chart">
 <div class="bar-row"><span class="bar-label">YOLO26</span><div class="bar-track"><div class="bar-fill bar-navy" style="width:78.0%"><span>31.21%</span></div></div></div>
 <div class="bar-row"><span class="bar-label">YOLO11</span><div class="bar-track"><div class="bar-fill bar-teal" style="width:71.9%"><span>28.75%</span></div></div></div>
 <div class="bar-row"><span class="bar-label">YOLOv8</span><div class="bar-track"><div class="bar-fill bar-navy" style="width:68.5%"><span>27.41%</span></div></div></div>
 </div>
<!-- ═══════════════════════════════════════════════════════ -->
 <!-- LEVEL 2: MODEL FAMILIES -->
 <!-- ═══════════════════════════════════════════════════════ -->
<!-- ── YOLO26 ──────────────────────────────────────────── -->
 <div class="family-section" id="yolo26">
 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <strong>YOLO26</strong></p>
 <h2>YOLO26</h2>
 <p>YOLO architecture with end-to-end attention head. <em>Note: <code>end2end=False</code> required for INT8 export.</em></p>
<h4>Tasks</h4>
 <div class="model-grid">
 <div class="model-card">
 <h3><a href="#yolo26-det" target="_self">Detection</a></h3>
 <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 mAP@0.5: 55.06% &middot; <a href="https://huggingface.co/EdgeFirst/yolo26-det">HF Repo</a></p>
 </div>
 <div class="model-card">
 <h3><a href="#yolo26-seg" target="_self">Segmentation</a></h3>
 <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 Mask mAP: 31.21% &middot; <a href="https://huggingface.co/EdgeFirst/yolo26-seg">HF Repo</a></p>
 </div>
 </div>
<!-- Level 3: YOLO26 Detection -->
 <div id="yolo26-det">
 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolo26" target="_self">YOLO26</a> &rsaquo; <strong>Detection</strong></p>
 <h3>YOLO26 Detection &mdash; ONNX FP32 reference</h3>
 <p>Accuracy ceiling per size, measured against COCO val2017 (5000 images, 80 classes) via the EdgeFirst Profiler + Validator pipeline.</p>
 <table class="data-table">
 <tr><th>Size</th><th>Params</th><th>GFLOPs</th><th>mAP@0.5</th><th>mAP@0.5-0.95</th><th>mAP@0.75</th></tr>
 <tr><td>Nano</td><td class="num">2.7M</td><td class="num">7.6</td><td class="num">55.06%</td><td class="num">39.71%</td><td class="num">42.87%</td></tr>
 <tr><td>Small</td><td class="num">10.3M</td><td class="num">27.0</td><td class="num">63.60%</td><td class="num">47.16%</td><td class="num">51.14%</td></tr>
 <tr><td>Medium</td><td class="num">24.5M</td><td class="num">74.4</td><td class="num">68.89%</td><td class="num">51.88%</td><td class="num">56.41%</td></tr>
 </table>
 <div class="drill-down">
 <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
 <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolo26-det">View YOLO26 Detection model card on HuggingFace &rarr;</a>
 </div>
 </div>
<!-- Level 3: YOLO26 Segmentation -->
 <div id="yolo26-seg">
 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolo26" target="_self">YOLO26</a> &rsaquo; <strong>Segmentation</strong></p>
 <h3>YOLO26 Segmentation &mdash; ONNX FP32 reference</h3>
 <p>Box and mask accuracy ceiling per size. Split-decoder architecture; mask coefficients and prototype outputs decoded by the EdgeFirst HAL after dequantisation.</p>
 <table class="data-table">
 <tr><th>Size</th><th>Box mAP@0.5</th><th>Box mAP@0.5-0.95</th><th>Mask mAP@0.5</th><th>Mask mAP@0.5-0.95</th></tr>
 <tr><td>Nano</td><td class="num">54.85%</td><td class="num">39.31%</td><td class="num">50.77%</td><td class="num">31.21%</td></tr>
 <tr><td>Small</td><td class="num">63.15%</td><td class="num">46.66%</td><td class="num">58.33%</td><td class="num">36.07%</td></tr>
 <tr><td>Medium</td><td class="num">68.80%</td><td class="num">51.76%</td><td class="num">63.36%</td><td class="num">39.47%</td></tr>
 </table>
 <div class="drill-down">
 <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
 <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolo26-seg">View YOLO26 Segmentation model card on HuggingFace &rarr;</a>
 </div>
 </div>
 <a class="back-to-top" href="#top" target="_self">&uarr; Back to top</a>
 </div>
<!-- ── YOLO11 ──────────────────────────────────────────── -->
 <div class="family-section" id="yolo11">
 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <strong>YOLO11</strong></p>
 <h2>YOLO11</h2>
 <p>Architecture with C3k2 attention blocks. Anchor-free DFL detection head.</p>
<h4>Tasks</h4>
 <div class="model-grid">
 <div class="model-card">
 <h3><a href="#yolo11-det" target="_self">Detection</a></h3>
 <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 mAP@0.5: 53.05% &middot; <a href="https://huggingface.co/EdgeFirst/yolo11-det">HF Repo</a></p>
 </div>
 <div class="model-card">
 <h3><a href="#yolo11-seg" target="_self">Segmentation</a></h3>
 <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 Mask mAP: 28.75% &middot; <a href="https://huggingface.co/EdgeFirst/yolo11-seg">HF Repo</a></p>
 </div>
 </div>
<div id="yolo11-det">
 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolo11" target="_self">YOLO11</a> &rsaquo; <strong>Detection</strong></p>
 <h3>YOLO11 Detection &mdash; ONNX FP32 reference</h3>
 <p>Accuracy ceiling per size, measured against COCO val2017 (5000 images, 80 classes) via the EdgeFirst Profiler + Validator pipeline.</p>
 <table class="data-table">
 <tr><th>Size</th><th>Params</th><th>GFLOPs</th><th>mAP@0.5</th><th>mAP@0.5-0.95</th><th>mAP@0.75</th></tr>
 <tr><td>Nano</td><td class="num">2.6M</td><td class="num">6.5</td><td class="num">53.05%</td><td class="num">37.83%</td><td class="num">41.07%</td></tr>
 <tr><td>Small</td><td class="num">9.4M</td><td class="num">21.5</td><td class="num">61.21%</td><td class="num">44.91%</td><td class="num">48.44%</td></tr>
 <tr><td>Medium</td><td class="num">20.1M</td><td class="num">68.0</td><td class="num">65.91%</td><td class="num">49.59%</td><td class="num">53.80%</td></tr>
 </table>
 <div class="drill-down">
 <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
 <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolo11-det">View YOLO11 Detection model card on HuggingFace &rarr;</a>
 </div>
 </div>
<div id="yolo11-seg">
 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolo11" target="_self">YOLO11</a> &rsaquo; <strong>Segmentation</strong></p>
 <h3>YOLO11 Segmentation &mdash; ONNX FP32 reference</h3>
 <p>Box and mask accuracy ceiling per size. Split-decoder architecture; mask coefficients and prototype outputs decoded by the EdgeFirst HAL after dequantisation.</p>
 <table class="data-table">
 <tr><th>Size</th><th>Box mAP@0.5</th><th>Box mAP@0.5-0.95</th><th>Mask mAP@0.5</th><th>Mask mAP@0.5-0.95</th></tr>
 <tr><td>Nano</td><td class="num">52.00%</td><td class="num">37.22%</td><td class="num">47.88%</td><td class="num">28.75%</td></tr>
 <tr><td>Small</td><td class="num">60.60%</td><td class="num">44.56%</td><td class="num">55.46%</td><td class="num">33.33%</td></tr>
 <tr><td>Medium</td><td class="num">65.91%</td><td class="num">49.59%</td><td class="num">60.54%</td><td class="num">36.61%</td></tr>
 </table>
 <div class="drill-down">
 <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
 <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolo11-seg">View YOLO11 Segmentation model card on HuggingFace &rarr;</a>
 </div>
 </div>
 <a class="back-to-top" href="#top" target="_self">&uarr; Back to top</a>
 </div>
<!-- ── YOLOv8 ──────────────────────────────────────────── -->
 <div class="family-section" id="yolov8">
 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <strong>YOLOv8</strong></p>
 <h2>YOLOv8</h2>
 <p>Anchor-free DFL detection head. Detection and instance-segmentation variants.</p>
<h4>Tasks</h4>
 <div class="model-grid">
 <div class="model-card">
 <h3><a href="#yolov8-det" target="_self">Detection</a></h3>
 <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 mAP@0.5: 50.55% &middot; <a href="https://huggingface.co/EdgeFirst/yolov8-det">HF Repo</a></p>
 </div>
 <div class="model-card">
 <h3><a href="#yolov8-seg" target="_self">Segmentation</a></h3>
 <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 Mask mAP: 27.41% &middot; <a href="https://huggingface.co/EdgeFirst/yolov8-seg">HF Repo</a></p>
 </div>
 </div>
<div id="yolov8-det">
 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolov8" target="_self">YOLOv8</a> &rsaquo; <strong>Detection</strong></p>
 <h3>YOLOv8 Detection &mdash; ONNX FP32 reference</h3>
 <p>Accuracy ceiling per size, measured against COCO val2017 (5000 images, 80 classes) via the EdgeFirst Profiler + Validator pipeline.</p>
 <table class="data-table">
 <tr><th>Size</th><th>Params</th><th>GFLOPs</th><th>mAP@0.5</th><th>mAP@0.5-0.95</th><th>mAP@0.75</th></tr>
 <tr><td>Nano</td><td class="num">3.2M</td><td class="num">8.9</td><td class="num">50.55%</td><td class="num">35.87%</td><td class="num">38.92%</td></tr>
 <tr><td>Small</td><td class="num">11.2M</td><td class="num">28.8</td><td class="num">59.39%</td><td class="num">43.23%</td><td class="num">46.69%</td></tr>
 <tr><td>Medium</td><td class="num">25.9M</td><td class="num">79.3</td><td class="num">64.70%</td><td class="num">48.46%</td><td class="num">52.62%</td></tr>
 </table>
 <div class="drill-down">
 <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
 <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolov8-det">View YOLOv8 Detection model card on HuggingFace &rarr;</a>
 </div>
 </div>
<div id="yolov8-seg">
 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolov8" target="_self">YOLOv8</a> &rsaquo; <strong>Segmentation</strong></p>
 <h3>YOLOv8 Segmentation &mdash; ONNX FP32 reference</h3>
 <p>Box and mask accuracy ceiling per size. Split-decoder architecture; mask coefficients and prototype outputs decoded by the EdgeFirst HAL after dequantisation.</p>
 <table class="data-table">
 <tr><th>Size</th><th>Box mAP@0.5</th><th>Box mAP@0.5-0.95</th><th>Mask mAP@0.5</th><th>Mask mAP@0.5-0.95</th></tr>
 <tr><td>Nano</td><td class="num">50.04%</td><td class="num">35.28%</td><td class="num">46.03%</td><td class="num">27.41%</td></tr>
 <tr><td>Small</td><td class="num">58.85%</td><td class="num">43.05%</td><td class="num">53.98%</td><td class="num">32.55%</td></tr>
 <tr><td>Medium</td><td class="num">63.86%</td><td class="num">47.84%</td><td class="num">58.74%</td><td class="num">35.59%</td></tr>
 </table>
 <div class="drill-down">
 <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
 <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolov8-seg">View YOLOv8 Segmentation model card on HuggingFace &rarr;</a>
 </div>
 </div>
 <a class="back-to-top" href="#top" target="_self">&uarr; Back to top</a>
 </div>
<!-- ── YOLOv5 ──────────────────────────────────────────── -->
 <div class="family-section" id="yolov5">
 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <strong>YOLOv5</strong></p>
 <h2>YOLOv5</h2>
 <p>CSP-Darknet backbone with anchor-based detection head. Detection only.</p>
<h4>Tasks</h4>
 <div class="model-grid">
 <div class="model-card">
 <h3><a href="#yolov5-det" target="_self">Detection</a></h3>
 <p class="meta">n/s/m sizes &middot; Nano ONNX FP32 mAP@0.5: 47.77% &middot; <a href="https://huggingface.co/EdgeFirst/yolov5-det">HF Repo</a></p>
 </div>
 </div>
<div id="yolov5-det">
 <p class="breadcrumb"><a href="#top" target="_self">Models</a> &rsaquo; <a href="#yolov5" target="_self">YOLOv5</a> &rsaquo; <strong>Detection</strong></p>
 <h3>YOLOv5 Detection &mdash; ONNX FP32 reference</h3>
 <p>Accuracy ceiling per size, measured against COCO val2017 (5000 images, 80 classes) via the EdgeFirst Profiler + Validator pipeline.</p>
 <table class="data-table">
 <tr><th>Size</th><th>Params</th><th>GFLOPs</th><th>mAP@0.5</th><th>mAP@0.5-0.95</th><th>mAP@0.75</th></tr>
 <tr><td>Nano</td><td class="num">1.9M</td><td class="num">4.5</td><td class="num">47.77%</td><td class="num">32.93%</td><td class="num">35.68%</td></tr>
 <tr><td>Small</td><td class="num">7.2M</td><td class="num">16.5</td><td class="num">57.32%</td><td class="num">41.27%</td><td class="num">44.90%</td></tr>
 <tr><td>Medium</td><td class="num">21.2M</td><td class="num">49.0</td><td class="num">63.29%</td><td class="num">47.00%</td><td class="num">51.39%</td></tr>
 </table>
 <div class="drill-down">
 <p><strong>Per-target validation and pipelined throughput</strong> &mdash; INT8 accuracy and profiler-measured pipelined FPS on i.MX 8M Plus, i.MX 95 Neutron, NXP Ara-240, RPi5 + Hailo-8L, and NVIDIA Jetson, each row linked to its EdgeFirst Studio session.</p>
 <a class="cta-link" href="https://huggingface.co/EdgeFirst/yolov5-det">View YOLOv5 Detection model card on HuggingFace &rarr;</a>
 </div>
 </div>
 <a class="back-to-top" href="#top" target="_self">&uarr; Back to top</a>
 </div>
<!-- ═══════════════════════════════════════════════════════ -->
 <!-- NAMING & VARIANTS -->
 <!-- ═══════════════════════════════════════════════════════ -->
<h2>Naming Convention</h2>
 <p>Each HuggingFace repo contains one model family for one task, organized by platform folders.</p>
 <table class="data-table">
 <tr><th>Component</th><th>Pattern</th><th>Example</th></tr>
 <tr><td>HF Repo</td><td><code>EdgeFirst/{version}-{task}</code></td><td><code>EdgeFirst/yolov8-det</code></td></tr>
 <tr><td>ONNX</td><td><code>{ver}{sz}-{task}-{prec}.onnx</code></td><td><code>yolov8n-det-fp32.onnx</code></td></tr>
 <tr><td>TFLite</td><td><code>{ver}{sz}-{task}-{prec}.tflite</code></td><td><code>yolov8n-det-int8.tflite</code></td></tr>
 <tr><td>Smart</td><td><code>{ver}{sz}-{task}-{prec}-smart.tflite</code></td><td><code>yolov8n-seg-int8-smart.tflite</code></td></tr>
 <tr><td>i.MX 95</td><td><code>{ver}{sz}-{task}-{prec}.imx95.tflite</code></td><td><code>yolov8n-det-int8.imx95.tflite</code></td></tr>
 <tr><td>Hailo HEF</td><td><code>{ver}{sz}-{task}-{prec}.hailo8l.hef</code></td><td><code>yolov8n-det-int8.hailo8l.hef</code></td></tr>
 <tr><td>TensorRT</td><td><code>{ver}{sz}-{task}-{prec}.{gpu}.engine</code></td><td><code>yolov8n-det-fp16.orin-nano.engine</code></td></tr>
 </table>
<h3>Decoder Variants</h3>
 <p>INT8 quantized models are available in two decoder configurations. The <strong>default</strong> (no suffix) uses a logical split-decoder &mdash; lossless and zero additional cost. The <strong>-smart</strong> variant uses a multi-scale split-decoder for improved accuracy at the cost of more CPU post-processing. Both are published so users can choose the tradeoff.</p>
<!-- ═══════════════════════════════════════════════════════ -->
 <!-- ROADMAP -->
 <!-- ═══════════════════════════════════════════════════════ -->
<h2>Roadmap</h2>
 <p>Expanding across the full spatial perception stack. All models validated on real hardware.</p>
 <table class="data-table">
 <tr><th>Category</th><th>Examples</th><th>Platforms</th><th>Status</th></tr>
 <tr>
 <td>Detection</td>
 <td>DETR-class, EfficientDet, mobile-optimized</td>
 <td><div class="badge-row"><span class="platform-badge">i.MX</span> <span class="platform-badge">Ara240</span> <span class="platform-badge">Hailo</span> <span class="platform-badge">Jetson</span></div></td>
 <td><span class="status-coming">Coming Soon</span></td>
 </tr>
 <tr>
 <td>Semantic Seg</td>
 <td>Lightweight real-time scene parsing</td>
 <td><div class="badge-row"><span class="platform-badge">i.MX</span> <span class="platform-badge">Ara240</span> <span class="platform-badge">Hailo</span> <span class="platform-badge">Jetson</span></div></td>
 <td><span class="status-planned">Roadmap</span></td>
 </tr>
 <tr>
 <td>SAM Seg</td>
 <td>Prompted, class-agnostic masks</td>
 <td><div class="badge-row"><span class="platform-badge">Ara240</span> <span class="platform-badge">Jetson</span></div></td>
 <td><span class="status-planned">Roadmap</span></td>
 </tr>
 <tr>
 <td>Monocular Depth</td>
 <td>Relative and metric depth estimation</td>
 <td><div class="badge-row"><span class="platform-badge">i.MX</span> <span class="platform-badge">Ara240</span> <span class="platform-badge">Jetson</span></div></td>
 <td><span class="status-planned">Roadmap</span></td>
 </tr>
 <tr>
 <td>3D &amp; Occupancy</td>
 <td>Monocular 3D, BEV, occupancy grids</td>
 <td><div class="badge-row"><span class="platform-badge">Jetson</span></div></td>
 <td><span class="status-planned">Roadmap</span></td>
 </tr>
 <tr>
 <td>Edge VLMs</td>
 <td>Visual language models for edge</td>
 <td><div class="badge-row"><span class="platform-badge">Ara240</span> <span class="platform-badge">Jetson</span></div></td>
 <td><span class="status-planned">Roadmap</span></td>
 </tr>
 </table>
 <p class="roadmap-note">Roadmap is subject to change. Models are published as validation completes on each target platform.</p>
<!-- ═══════════════════════════════════════════════════════ -->
 <!-- ARCHITECTURE & STUDIO -->
 <!-- ═══════════════════════════════════════════════════════ -->
<h2>EdgeFirst Studio</h2>
 <p>
 <a href="https://edgefirst.studio"><strong>EdgeFirst Studio</strong></a> is the MLOps platform that drives the entire model zoo pipeline. <strong>Free tier available.</strong>
 </p>
 <ul class="studio-features">
 <li>Dataset management &amp; AI-assisted annotation</li>
 <li>Model training with automatic multi-format export and INT8 quantization</li>
 <li>Reference and on-target validation with full metrics collection</li>
 <li>CameraAdaptor integration for native sensor format training</li>
 <li>Deploy trained models to edge devices via the <a href="https://github.com/EdgeFirstAI/client">EdgeFirst Client</a> CLI</li>
 </ul>
<div class="footer">
 <p>Apache 2.0 &middot; &copy; <a href="https://www.au-zone.com">Au-Zone Technologies Inc.</a></p>
 </div>
 </div>
</body>
</html>