<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by SharkYun on Medium]]></title>
        <description><![CDATA[Stories by SharkYun on Medium]]></description>
        <link>https://medium.com/@sharkyun?source=rss-2aa70868c3aa------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*I5JImkjlbMVowKIIrqILQg.jpeg</url>
            <title>Stories by SharkYun on Medium</title>
            <link>https://medium.com/@sharkyun?source=rss-2aa70868c3aa------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 07:35:21 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sharkyun/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[From Industrial AI to Publication: Second SCI Paper Accepted by Electronics]]></title>
            <link>https://sharkyun.medium.com/from-industrial-ai-to-publication-second-sci-paper-accepted-by-electronics-96081f726857?source=rss-2aa70868c3aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/96081f726857</guid>
            <dc:creator><![CDATA[SharkYun]]></dc:creator>
            <pubDate>Wed, 12 Aug 2026 05:46:46 GMT</pubDate>
            <atom:updated>2026-08-19T07:06:35.424Z</atom:updated>
            <content:encoded><![CDATA[<h3>Research Milestone: My Second SCI Paper Accepted in <em>Electronics</em> 🎉</h3><p>I am happy to share that our research paper has been accepted for publication in <strong>Electronics (MDPI)</strong> under the <strong>Artificial Intelligence</strong> section and the Special Issue <strong>“Applications of Image Analysis and Intelligent Vision.”</strong><a href="https://www.mdpi.com/journal/electronics/special_issues/17127O0LI8">Applications of Image Analysis and Intelligent Vision</a></p><h3>Paper Timeline</h3><ul><li>Submitted: June 24, 2026</li><li>Major Revision: July 21, 2026</li><li>Revised: July 27, 2026</li><li>Accepted: August 10, 2026</li><li>Published: 18 August 2026</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*q7WsS1on_JDG5GT185cE4w.png" /></figure><h3>About This Work</h3><p>This study focuses on an industrial computer vision problem in precision wire welding. We proposed a framework that combines:</p><ul><li>YOLO-based conductor localization</li><li>EfficientViT-SAM-based segmentation</li><li>Geometric offset generation for alignment support</li></ul><p>The objective is not only to detect conductors accurately but also to transform visual inspection results into actionable alignment information that can support manufacturing operations.</p><h3>What I Learned</h3><p>One of the most interesting lessons from this project was that achieving high detection accuracy is only part of the challenge.</p><p>In real manufacturing environments, the key question is often:</p><blockquote><em>“How can model outputs be translated into meaningful engineering decisions?”</em></blockquote><p>This work pushed me to think beyond traditional object detection metrics and focus on how AI-generated visual information can directly support process improvement and production quality.</p><h3>Personal Reflection</h3><p>I am especially grateful that this work has become <strong>my second SCI publication</strong> while working full-time as a data scientist and pursuing an online master’s degree.</p><p>Research is often a long journey filled with experiments, revisions, unexpected challenges, and continuous learning. This acceptance is not only a recognition of the work itself, but also a reminder that persistence and curiosity can gradually lead to meaningful outcomes.</p><p>I sincerely appreciate the support and guidance from my supervisors, collaborators, and research partners throughout this project.</p><p>Looking forward to continuing my journey in Industrial AI, Computer Vision, and Smart Manufacturing.</p><p>#ArtificialIntelligence #ComputerVision #DeepLearning #SmartManufacturing #IndustrialAI #MachineLearning #YOLO #SAM #Electronics #Research #DataScience</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=96081f726857" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My First SCI Paper Accepted by IJAMT: A 60-Day Journey from Submission to Acceptance]]></title>
            <link>https://sharkyun.medium.com/my-first-sci-paper-accepted-by-ijamt-a-60-day-journey-from-submission-to-acceptance-83599e5ce04c?source=rss-2aa70868c3aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/83599e5ce04c</guid>
            <dc:creator><![CDATA[SharkYun]]></dc:creator>
            <pubDate>Tue, 28 Jul 2026 01:51:41 GMT</pubDate>
            <atom:updated>2026-07-28T01:51:41.823Z</atom:updated>
            <content:encoded><![CDATA[<p>As a data scientist in the manufacturing industry, my role is split between industrial AI applications and research. While I regularly work on machine learning and computer vision projects, I never expected that one of those projects would eventually lead to my first SCI journal publication, especially as a first author while pursuing an online master’s degree.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zWq_MoAXOzGmrPmKT8D19w.png" /></figure><p>After months of research, experiments, writing, and revisions, I am excited to share that my first SCI journal paper has been accepted by <em>The International Journal of Advanced Manufacturing Technology (IJAMT)</em> as the <strong>first author</strong>.</p><p>As a newcomer to academic publishing, I spent a lot of time searching online for real submission timelines and review experiences. Therefore, I would like to share my own journey and hopefully give some reference to researchers who are preparing their first journal submission.</p><h3>Overall Statistics</h3><ul><li>Submission to first decision: <strong>53 days</strong></li><li>Revision period: <strong>2 days</strong></li><li>Revision submission to acceptance: <strong>4 days</strong></li><li>Total submission to acceptance: <strong>60 days</strong></li></ul><p><strong>Timeline Summary</strong></p><p>📅 May 25, 2026 → Submitted<br> 📅 May 27, 2026 → Missing data request<br> 📅 May 28 — July 17, 2026 → Under Review<br> 📅 July 17, 2026 → Minor Revision<br> 📅 July 19, 2026 → Revision Submitted<br> 📅 July 23, 2026 → Accepted</p><h3>What I Learned</h3><h3>1. The first submission doesn’t need to be perfect</h3><p>Before submitting, I constantly worried that the manuscript was not good enough. Looking back, many aspects were indeed improved during peer review. Reviewers helped identify weaknesses that we had overlooked.</p><h3>2. Reviewer comments are opportunities</h3><p>At first, reviewer feedback can feel intimidating. However, most comments genuinely aim to improve the quality and clarity of the work. Responding carefully often makes the paper significantly better.</p><h3>3. Reference management matters</h3><p>A surprising amount of revision time was spent updating references, correcting citations, and ensuring consistency throughout the manuscript.</p><h3>4. Patience is part of the process</h3><p>Research teaches technical skills, but publishing teaches patience. The waiting period can be challenging, especially for a first submission.</p><h3>Final Thoughts</h3><p>This acceptance represents an important milestone in my research journey: <strong>my first published SCI paper and my first first-author publication</strong>.</p><p>I am deeply grateful to my supervisor, collaborators, reviewers, and everyone who contributed throughout the process.</p><p>For anyone currently waiting for reviewer comments, revising a manuscript, or preparing a first submission: keep going. The journey may feel long, but the acceptance email is worth the wait.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=83599e5ce04c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[World Models Are Coming — The Next Phase of AI?]]></title>
            <link>https://sharkyun.medium.com/world-models-are-coming-the-next-phase-of-ai-5c67807f0f90?source=rss-2aa70868c3aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/5c67807f0f90</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[world-models]]></category>
            <dc:creator><![CDATA[SharkYun]]></dc:creator>
            <pubDate>Fri, 20 Mar 2026 01:30:37 GMT</pubDate>
            <atom:updated>2026-03-20T01:31:01.586Z</atom:updated>
            <content:encoded><![CDATA[<p>Last week, Reuters reported that <strong>Yann LeCun</strong>’s Advanced Machine Intelligence (<strong>AMI</strong>) raised<strong> $1.03B</strong> to pursue an alternative path in AI — one that doesn’t rely purely on scaling ever-larger language models.</p><p>For years, progress in AI has been dominated by scaling laws: bigger models, more data, more compute. Large Language Models proved something important: fluent behavior and broad capabilities can <em>emerge</em> from scale.</p><p>But even as LLMs became impressively articulate, a fundamental limitation stayed visible:</p><ul><li>They don’t truly understand the world they talk about</li><li>Much of what they learn is correlation, not causation</li><li>They’re exceptional pattern recognizers, but often weak at grounded reasoning — especially about physical reality, space, time, and action</li></ul><p>That gap is exactly where World Models enter the conversation.</p><h3>Yann LeCun and the JEPA Vision</h3><p>Yann LeCun, who led Meta AI (FAIR) for many years, has been increasingly vocal about a core belief:</p><blockquote><strong><em>Generative models that predict pixels or tokens are not enough.</em></strong></blockquote><p>Instead, he argues that intelligence should be built on <strong>predictive internal representations</strong>, not raw reconstructions. That philosophy crystallized into <strong>JEPA — Joint Embedding Predictive Architecture</strong>.</p><p>In his view, intelligence should not be built on raw reconstruction — whether pixels, words, or frames — but on <strong>predictive internal representations</strong> that capture how the world <em>works</em>.</p><p>That philosophy crystallized into <strong>JEPA — Joint Embedding Predictive Architecture</strong>.</p><p>Instead of predicting pixels or tokens directly, JEPA:</p><ul><li>Learns <strong>representations (embeddings)</strong> of observations</li><li>Predicts <strong>missing or future information in representation space</strong></li><li>Avoids modeling irrelevant details like texture noise or exact appearance</li></ul><p>The goal is to learn <em>what matters</em>:</p><ul><li>objects</li><li>structure</li><li>dynamics</li><li>causality</li></ul><p>Without being forced to regenerate every visual or lexical detail.</p><p>In LeCun’s framing:</p><blockquote><strong><em>Don’t ask the model to imagine what the world looks like — ask it to understand what the world is.</em></strong></blockquote><h3>JEPA 2.1: When World Models Meet Dense Vision</h3><p>Early JEPA and V‑JEPA models were strong at <strong>global reasoning</strong>:</p><ul><li>Action recognition</li><li>Temporal understanding</li><li>Planning and control</li></ul><p>But they also had a clear weakness.</p><p>They were <strong>not great vision backbones</strong>.</p><p>Specifically:</p><ul><li>Dense features were weak</li><li>Spatial grounding was fragile</li><li>Local structure was hard to extract</li></ul><p>This limited their usefulness for tasks like segmentation, depth estimation, and tracking — areas where models like DINO excelled.</p><h3>Turning Point: JEPA 2.1</h3><p>[<a href="https://arxiv.org/abs/2603.14482">paper</a>]</p><p>JEPA 2.1 feels less like an incremental update and more like a <strong>fusion</strong>:</p><blockquote><strong><em>JEPA + DINO‑style dense representation learning</em></strong></blockquote><p>By redesigning how prediction loss is applied and how spatial structure is supervised, JEPA 2.1 manages to:</p><ul><li>Preserve <strong>world‑model strengths</strong> (prediction, dynamics, causality)</li><li>While unlocking <strong>high‑quality dense visual features</strong></li></ul><p>For the first time, a JEPA‑based model can plausibly act as:</p><ul><li>A strong vision backbone</li><li>A temporal reasoning system</li><li>A foundation for planning and embodied intelligence</li></ul><h3>A Bridge</h3><p>JEPA 2.1 shouldn’t be seen as the destination. It’s a <strong>bridge</strong>:</p><ul><li>From static perception → dynamic understanding</li><li>From representation learning → world modeling</li><li>From task‑specific heads → general planning systems</li></ul><h3>Why This Matters Now</h3><p>The renewed focus on world models — and the scale of investment behind efforts like AMI — suggests a growing consensus:</p><blockquote><em>Scaling alone may not be enough.</em></blockquote><p>To build systems that can reason, plan, and act in the real world, AI may need:</p><ul><li>Predictive internal states</li><li>Causal structure</li><li>Temporal consistency</li><li>Physical grounding</li></ul><p>This might be the next phase of AI~</p><p>News source (Reuters):<br><a href="https://www.reuters.com/business/ex-meta-ai-chief-yann-lecuns-ami-raises-103-billion-alternative-ai-approach-2026-03-10/">https://www.reuters.com/business/ex-meta-ai-chief-yann-lecuns-ami-raises-103-billion-alternative-ai-approach-2026-03-10/</a></p><p>JEPA [<a href="https://arxiv.org/abs/2301.08243">paper</a>]</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5c67807f0f90" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Computer Vision in AI: From Pixels to Perception]]></title>
            <link>https://sharkyun.medium.com/computer-vision-in-ai-from-pixels-to-perception-d233a1914b3a?source=rss-2aa70868c3aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/d233a1914b3a</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[SharkYun]]></dc:creator>
            <pubDate>Wed, 17 Dec 2025 08:07:11 GMT</pubDate>
            <atom:updated>2025-12-17T08:07:11.376Z</atom:updated>
            <content:encoded><![CDATA[<p>When we talk about AI, AI, AI blah blah blah. Its kinda hard to people to realte what Computer Vision (<strong>CV</strong>) as a field in AI. What Is Computer Vision? Where It Fits in AI and the 4 Core Tasks. (Classification, Detection, Segmentation, Pose)</p><blockquote><strong>Computer Vision</strong> is the part of AI that gives machines the ability to <strong>“see”</strong> and understand visual data (images, video, 3D), so they can measure, describe, and act on the physical world.</blockquote><p>As we know <strong>AI </strong>have <strong>ML </strong>and <strong>DL</strong>, and <strong>CV </strong>is simply a subfield of AI and larget intergal with image/videos by camera. Computer Vision is the eyes of AI: it turns pixels from cameras and sensors into facts about the world — what’s there, where it is, how it’s moving — so other AI components can make decision</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/591/0*mFFdSxXfklrqm-18.png" /><figcaption>AI sub-fields <a href="https://ar.inspiredpencil.com/pictures-2023/artificial-intelligence-computer-vision">Source</a></figcaption></figure><h3>What Computer Vision Does.</h3><h4>1. Classification</h4><p>Identifying what types of objects are present in an image (e.g., cars, people, animals).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JuZscVxFulwGnwe4.jpg" /><figcaption><a href="https://imagevision.ai/blog/understanding-computer-vision-a-technical-overview/">Understanding Computer Vision | ImageVision.ai</a></figcaption></figure><h4>2. Object Detection</h4><p>= <strong>Classification+Localization</strong></p><ol><li><strong>Classification</strong>: Identifying what types of objects are present in an image (e.g., cars, people, animals).</li><li><strong>Localization</strong>: Determining where these objects are located within the image by drawing bounding boxes around them.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*NgW301IdC-pszjuK.png" /><figcaption>by SharkYun</figcaption></figure><p>More Detail at <a href="https://medium.com/@sharkyun/computer-vision-object-detection-one-stage-vs-two-stage-b05dbff88195">HERE</a></p><h4>3. Segmentation</h4><p>Idea: Predict a label for every pixel (mask). You can think of it as pixel-wise classification.</p><ul><li><strong>Semantic </strong>segmentation: Assign a “stuff”class to each pixel (road, sky, building).</li><li><strong>Instance </strong>segmentation: Separate each object as “thing” instance and give each a mask + class.</li><li><strong>Panoptic </strong>segmentation:Unified view — “thing” classes as instances + “stuff” classes as semantic regions.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*4ry0YqRJ7Ih8JzE8feQ_kg.png" /><figcaption>Source from META</figcaption></figure><h4>4. Keypoints &amp; Pose Estimation</h4><ul><li>Goal: Localize landmarks and estimate pose.</li><li>Output: 2D/3D keypoints or skeletal joints (face/hand/human/body pose).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/626/1*_NVExBjrcqbjC7I1CtuphQ.png" /></figure><p><a href="https://developers.google.com/ml-kit/vision/pose-detection">Pose detection | ML Kit | Google for Developers</a></p><h4>Combining these models, we use in many Cases</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QYiYQFWk0JCMiiC6.png" /></figure><p>Heres a link by google, it has demonsactaion many use case in CV</p><p><a href="https://developers.google.com/ml-kit/vision">https://developers.google.com/ml-kit/vision</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/204/1*MiIKSq4afXt1NJ_hyj8iHQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d233a1914b3a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Passed DP-100 with Minimal Preparation]]></title>
            <link>https://sharkyun.medium.com/how-i-passed-dp-100-with-minimal-preparation-276528f98d89?source=rss-2aa70868c3aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/276528f98d89</guid>
            <category><![CDATA[azure]]></category>
            <category><![CDATA[dp-100-exam]]></category>
            <dc:creator><![CDATA[SharkYun]]></dc:creator>
            <pubDate>Sat, 28 Jun 2025 08:20:40 GMT</pubDate>
            <atom:updated>2025-06-28T08:20:57.703Z</atom:updated>
            <content:encoded><![CDATA[<p>I recently took the DP-100 exam after completing AI-900, DP-900, and AZ-900 within a month. I took a short break before starting my DP-100 preparation, which lasted about two weeks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hlw24KasT4YsCXVfyptkqA.png" /></figure><blockquote>Status: <strong>Pass</strong></blockquote><blockquote>My score: <strong>700/1000</strong></blockquote><p>During the first 20 questions, I honestly thought I was going to fail. Many questions were quite different from the practice sets I used. My preparation was mainly based on:</p><ul><li>Microsoft Learn modules and their practice sets</li><li>O’Reilly’s practice questions</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xbryNWNdZoYsM5ZRxB6YTg.png" /></figure><p>However, I realized the exam was updated on <strong>April 11, 2025</strong>, and my practice sets were likely from an older version. The newer questions focused more on <strong>AI Foundry topics</strong> and how to evaluate LLMs, which were not covered in detail in my preparation materials.</p><p>In the end, I was lucky to pass with exactly 700. 😅</p><p>Just sharing this so that others preparing for DP-100 can <strong>check for the latest question trends and updates</strong>, especially around LLM evaluation and AI solution design.</p><p>Wish you all the best in your preparation and exams!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=276528f98d89" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Best Value Top-Tier Online CS Master’s Programs in 2025: Flexibility & Affordability]]></title>
            <link>https://sharkyun.medium.com/best-value-top-tier-online-cs-masters-programs-in-2025-flexibility-affordability-28776747cba9?source=rss-2aa70868c3aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/28776747cba9</guid>
            <category><![CDATA[best-value]]></category>
            <category><![CDATA[best-school]]></category>
            <category><![CDATA[graduate-school]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[top-10]]></category>
            <dc:creator><![CDATA[SharkYun]]></dc:creator>
            <pubDate>Sun, 23 Mar 2025 07:37:56 GMT</pubDate>
            <atom:updated>2025-06-09T07:12:40.212Z</atom:updated>
            <content:encoded><![CDATA[<p>In 2025, many institutions offer online Master Program we can pick from. Even the TOP tier CS school like <strong>Stanford, CMU, UC Berkeley, Georgia Tech </strong>does have online CS master. With so many options available, the question becomes: how do we choose the right program — assuming we get accepted?</p><h3>My Choice: GA Tech &amp; UT Austin</h3><p>I ultimately applied to the <strong>Online Master of Science in Computer Science (OMSCS) at Georgia Institute of Technology</strong> and the <strong>Master of Science in Artificial Intelligence (MSAI) at the University of Texas at Austin</strong>. My decision was based on a combination of flexibility, affordability, prestige, and academic quality.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ayqlTwIL2AqFBmUkCmgzYA.png" /></figure><h3>Flexibility</h3><p>A major reason for choosing an online program was the ability to <strong>pursue my master’s while continuing to work</strong>. The flexible scheduling allows me to balance professional responsibilities with advanced studies, making it an ideal option for working professionals. While I recognize that applying on-site might increase my chances of admission to an even more prestigious program, the higher cost and potential sacrifice of valuable family time made the online option more appealing.</p><p>Furthermore, OMSCS offers <strong>six different specialization paths</strong> that you can choose from later, whereas UT Austin’s College of Natural Sciences (CDSO) requires you to select one of three tracks — AI, CS, or DS — at the time of admission.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JicZJo3a6ThkPJP9je-JBA.png" /><figcaption>GA Tech OMSCS offer 6 path</figcaption></figure><h3>Best Value</h3><p>The primary factor in my choice was affordability. These programs cost <strong>around $10,000 for a two-year degree</strong>, which is incredibly reasonable for institutions of this caliber. In contrast, many other top CS programs charge anywhere between <strong>$30,000 to $80,000 or more</strong>.</p><h4>TOP-tier CS School</h4><p>Base on <a href="https://www.usnews.com/best-graduate-schools/top-science-schools/artificial-intelligence-rankings">USNews</a>, they both surely are in TOP10 school in CS/AI field. World Ranking from Times are both in TOP50.</p><blockquote><strong>GTech</strong> : World Rank <strong>#40</strong>, #7 in CS, #5 in AI</blockquote><blockquote><strong>UTexas</strong>: World Rank <strong>#50</strong>, #7 in CS, #9 in AI</blockquote><p><a href="https://www.usnews.com/best-graduate-schools/top-science-schools/artificial-intelligence-rankings">The Best Artificial Intelligence Programs in America, Ranked</a></p><p>These rankings highlight the strength of their research and coursework, making them competitive choices for anyone looking to advance in the field of AI and computer science.</p><h3>The Best Value for Data Science and AI: for me</h3><p>For those seeking the best return on investment, <strong>OMSCS</strong> and <strong>MSAI</strong> stand out as some of the most cost-effective, high-quality programs available. With strong curricula, respected faculty, affordability, and the flexibility to work while studying and maintain a healthy family life, they offer a compelling alternative to traditional, high-cost graduate programs.</p><h3>Conclusion</h3><p>There’s never a single “best choice” because one size doesn’t fit all. Choose your school wisely and always believe in the path you take. Cheers!</p><h4>Reference</h4><p><a href="https://www.mastersinai.org/degrees/online-masters-in-artificial-intelligence/">21 Best Online Master’s in Artificial Intelligence Programs</a></p><p><strong>Follow-Up Update:</strong><br>✅ Offer from Georgia Tech on March 03,</p><p>✅ Offer from UT Austin on June 05.</p><p>Feel free to ask question~</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=28776747cba9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Passed the DP-900 Exam in 7 Days Right After AI-900 Back to Back]]></title>
            <link>https://sharkyun.medium.com/how-i-passed-the-dp-900-exam-in-7-days-right-after-ai-900-back-to-back-3815b6765109?source=rss-2aa70868c3aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/3815b6765109</guid>
            <category><![CDATA[dp-900-exam]]></category>
            <category><![CDATA[azure-devops]]></category>
            <category><![CDATA[azure]]></category>
            <category><![CDATA[dp-900]]></category>
            <dc:creator><![CDATA[SharkYun]]></dc:creator>
            <pubDate>Mon, 17 Mar 2025 06:55:52 GMT</pubDate>
            <atom:updated>2025-03-17T06:55:52.693Z</atom:updated>
            <content:encoded><![CDATA[<p>After successfully earning my <strong>AI-900</strong> Azure AI Fundamentals certification, I decided to take on the <strong>DP-900</strong> Azure Data Fundamentals exam. Initially, I planned to take it with 2 week gap with AI-900, but due to some scheduling conflicts, I decided to take it earlier. It turned out to be a crazy yet rewarding week!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qwxO_QdQALe17HgHdQkPzQ.png" /><figcaption>Back to Back</figcaption></figure><h3>My Study Strategy: Intensive Learning and Practice</h3><p>The key to my success was a combination of focused video learning and extensive practice with sample questions. Here’s how I managed to prepare effectively in such a short time:</p><h3>1. Accelerated Video Learning</h3><p>I utilized a comprehensive YouTube course to cover all the exam topics efficiently:</p><p>YouTube Video Course:</p><ul><li><a href="https://www.youtube.com/watch?v=V7v-Ri9P5F0">Azure Data Fundamentals Certification (DP-900) — Full Course to PASS the Exam</a></li></ul><p>To make the most of my time, I watched the videos at <strong>1.25x speed</strong>, which helped me absorb the material more quickly without missing important details.</p><h3>2. Back-to-Back Practice Tests</h3><h4>Microsoft Official Practice Assessment</h4><ul><li><a href="https://learn.microsoft.com/">Practice Assessment</a></li></ul><h4>GitHub Practice Test</h4><ul><li><a href="https://github.com/PacktPublishing/DP-900-Microsoft-Azure-Data-Fundamentals-Video-Course-Ques/blob/main/V18684_Practice%20Test%201.pdf">DP-900 Practice Test 1</a></li><li><a href="https://github.com/PacktPublishing/DP-900-Microsoft-Azure-Data-Fundamentals-Video-Course-Ques/blob/main/V18684_Practice%20Test%202.pdf">DP-900-Practice Test 2</a></li><li><a href="https://github.com/PacktPublishing/DP-900-Microsoft-Azure-Data-Fundamentals-Video-Course-Ques/blob/main/V18684_QuestionBank.pdf">DP-900-QuestionBank</a> (150 quesitons)</li></ul><h4>Sample Questions</h4><iframe src="https://drive.google.com/viewerng/viewer?url=https%3A//arch-center.azureedge.net/Learning/Credentials/DP-900-SampleQuestions.pdf&amp;embedded=true" width="600" height="780" frameborder="0" scrolling="no"><a href="https://medium.com/media/80bf3eab1338a99ce24f2a7ba6939b16/href">https://medium.com/media/80bf3eab1338a99ce24f2a7ba6939b16/href</a></iframe><h3>Cheat Sheet of Key Concepts</h3><h4>Main Structure (Very important)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gMU0kRKGdDkvh5qL2AXPqQ.png" /></figure><h4><strong>Cheat Sheet (random note)</strong></h4><blockquote><strong>Data Factory</strong> is used to run ETL pipelines. (without code)</blockquote><blockquote><strong>Databricks</strong> is used for processing large amounts of data, which is supported by multiple cloud providers.</blockquote><blockquote><strong>HDInsight</strong> is used to process large amounts of data by using Hadoop.</blockquote><blockquote><strong>Azure Synapse Analytics</strong> is an Azure native service built on Apache Spark.</blockquote><blockquote><strong>Azure table storage </strong>RowKey is unique within a partition, not within a table. Items in the same partitions are stored in a row key order.Tables cannot have indexes to speed up queries.</blockquote><blockquote><strong>Data lakes</strong> store files. Data Lake Storage Gen2 is used for storing huge amounts of data to be processed by services such as Databricks, Azure Synapse Analytics, and HDInsight</blockquote><blockquote><strong>Data warehouses</strong> use fact and dimension tables in a star/snowflake schema.</blockquote><blockquote><strong>Cubes</strong> are generated from a data warehouse but are a table themselves.</blockquote><blockquote><strong>Azure Files</strong> is used to store and share files by using SMB（Server Message Block）and NFS（Network File System）</blockquote><blockquote><strong>Cosmos</strong></blockquote><blockquote>The <strong>Cassandra</strong> API is used for tabular data in a column-family storage.</blockquote><blockquote>queried by using SQL.</blockquote><blockquote>The <strong>Gremlin</strong> API is used for graph databases. Qureied by graph</blockquote><blockquote><strong>MongoDB</strong> API stores data in the Binary JSON (BSON) format.</blockquote><blockquote>queried by using MongoDB Query</blockquote><blockquote><strong>Table</strong> API queried by using OData and LINQ queries.</blockquote><blockquote>Stream</blockquote><blockquote><strong>Azure IoT Hub and Azure Event Hubs</strong> can be used as sources for stream processing. -&gt; <strong>Data Lake Storage</strong></blockquote><blockquote><strong>Delta Lake</strong> is an open-source storage layer that adds support to Data Lake Storage for transactional consistency.</blockquote><blockquote><strong>Stream Analytics</strong> can handle stream processing from Kafka to Data Lake</blockquote><blockquote>Normalization uses the least possible amount of storage. Normalization optimizes for updates, inserts, and deletes.</blockquote><blockquote><strong>Azure SQL Database </strong>is a serverless platform as a service (PaaS) SQL instance. allows you to provision a single database on a dedicated server and has on-demand scalability. is the best option for create, read, update, and delete (CRUD) operations and uses the <strong>least amount</strong> of storage space,</blockquote><blockquote><strong>Azure SQL Managed Instance</strong>: fully mananged relational db, and 100% with MS SQL Server</blockquote><blockquote><strong>PostgreSQL</strong> is a hybrid relational-object database. You can store data in relational tables, or you can store custom data types with non-relational properties.</blockquote><blockquote><strong>MariaDB</strong> has built-in support for temporal data. It enables applications to query data as the data appeared in previous points in time.</blockquote><blockquote><strong>ORC</strong> and <strong>Parquet</strong> are columnar data formats.</blockquote><blockquote><strong>Avro</strong> is row-based</blockquote><blockquote><strong>CSV</strong> is used for delimited data.</blockquote><blockquote><strong>Analytical data workloads</strong> are highly <strong>denormalized</strong> and are optimized for <strong>read</strong> operations.</blockquote><blockquote><strong>Graph workloads</strong> store hierarchical data.</blockquote><blockquote><strong>Transactional workloads</strong> are optimized for create, read, update, and delete (<strong>CRUD</strong>) operations. (rely on relationship between entitiy) (Live applications)</blockquote><blockquote>Page blobs are optimized for random access and used for VHDs.</blockquote><blockquote>Append blobs cannot be updated.</blockquote><blockquote><strong>Block blobs</strong>. A block blob is handled as a set of blocks. Each block can vary in size, up to 4000 MiB. A block blob can contain up to 190.7 TiB (4000 MiB X 50,000 blocks), giving a maximum size of over 5000 MiB. The block is the smallest amount of data that can be read or written as an individual unit. Block blobs are best used to store discrete, large, binary objects that change infrequently.</blockquote><blockquote><strong>Page blobs</strong>. A page blob is organized as a collection of fixed size 512-byte pages. <strong>A page blob is optimized to support random read and write operations</strong>; you can fetch and store data for a single page if necessary. A page blob can hold up to 8 TB of data. Azure uses page blobs to implement virtual disk storage for virtual machines. ( are optimized for random access and used for <strong>VHDs</strong>. )</blockquote><blockquote><strong>Append blobs.</strong> An append blob is a block blob optimized to support append operations. You can only add blocks to the end of an append blob; updating or deleting existing blocks isn&#39;t supported. Each block can vary in size, up to 4 MB. The maximum size of an append blob is just over 195 GB.</blockquote><h3>Final Thoughts</h3><p>Passing the DP-900 exam in just seven days after the AI-900 was challenging but achievable with dedication and the right approach. If you’re considering taking the DP-900 exam soon after the AI-900, I hope my experience provides a helpful roadmap.</p><p>That said, my approach might work best for those who already have some knowledge of databases. It truly required a lot of effort and relentless focus. In the last two days, I tackled about 150–200 practice questions and thoroughly reviewed the answers. Most importantly, I took detailed <strong>notes on the questions I got wrong</strong>, which helped me understand my mistakes and avoid repeating them.</p><p>Remember, consistency is key. Stay focused, make the best use of your available resources, and believe in your ability to succeed.</p><p>Good luck on your certification journey!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3815b6765109" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I passed the AI-900 Exam: Personal Preparation Guide]]></title>
            <link>https://sharkyun.medium.com/how-i-passed-the-ai-900-exam-personal-preparation-guide-d35a776eb9a3?source=rss-2aa70868c3aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/d35a776eb9a3</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-900]]></category>
            <category><![CDATA[ai-900-exam]]></category>
            <category><![CDATA[azure-certification]]></category>
            <category><![CDATA[azure]]></category>
            <dc:creator><![CDATA[SharkYun]]></dc:creator>
            <pubDate>Thu, 06 Mar 2025 08:23:15 GMT</pubDate>
            <atom:updated>2025-03-07T08:10:03.662Z</atom:updated>
            <content:encoded><![CDATA[<h3>How I passed the AI-900 Exam <em>in 10 Days</em>: Personal Study Guide</h3><p><em>A Step-by-Step Guide to Help You Succeed in Azure AI-900 in 10 Days</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cYzR1t6LOSuVcP0vHPC84Q.png" /><figcaption>Generated by DALLE</figcaption></figure><p>Recently, I just passed the <strong>AI-900: Azure AI Fundamentals</strong> exam. Since it was paid for by my company, I thought, why not? It’s my first Azure certification, and here’s how I prepared for the exam.</p><h3>Exam Domains</h3><h4>AI-900 covers the following domains:</h4><ul><li>Describe Artificial Intelligence workloads and considerations (15–20%)</li><li>Describe fundamental principles of machine learning on Azure (20–25%)</li><li>Describe features of computer vision workloads on Azure (15–20%)</li><li>Describe features of Natural Language Processing (NLP) workloads on Azure (15–20%)</li><li>Describe features of generative AI workloads on Azure (15–20%)</li></ul><h3>Preparation Steps</h3><h4>1. Start with Microsoft’s Free Learning Paths</h4><p>Microsoft offers free, self-paced learning modules tailored for the AI-900 exam.</p><ul><li>Access the Learning Paths: <a href="https://docs.microsoft.com/en-us/learn/certifications/azure-ai-fundamentals/">Microsoft Learn — AI Fundamentals</a></li></ul><p>The modules include:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/814/1*l3vR0jhED5MSUXoPhg_XTA.png" /></figure><ul><li>Explore the basics of AI</li><li>Introduction to Machine Learning</li><li>Build machine learning models</li><li>Introduction to Computer Vision</li><li>Explore Natural Language Processing</li><li>Introduction to Conversational AI</li></ul><blockquote>Tip: Even if you’re familiar with the content, complete the <strong>Knowledge Checks</strong> at the end of each module to assess your understanding.</blockquote><h4>2. Practice with Sample Questions</h4><h4>Free Resources</h4><ul><li><a href="https://docs.microsoft.com/en-us/learn/certifications/azure-ai-fundamentals/">Microsoft Learn — AI Fundamentals</a></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/273/1*vMxaH1IvVU3aegSlrOSghQ.png" /></figure><ul><li><a href="https://query.prod.cms.rt.microsoft.com/cms/api/am/binary/RWSHpv">Free Sample Questions</a>: Utilize any free practice questions available online to get a feel for the exam format.</li><li><a href="https://learn.microsoft.com/en-us/credentials/certifications/resources/study-guides/ai-900">Official Study Guide</a>: Microsoft’s official study guide outlines all the topics and provides insights into what areas to focus on.</li></ul><h4>Paid Resources</h4><ul><li>AI-900: Microsoft Azure AI Fundamentals (Pearson Practice Test): This practice test offers pretty similar testset as real one.</li><li>AI-900 Microsoft Azure AI Fundamentals Certification Guide by Julian Sharp</li></ul><h3>Exam Tips</h3><ul><li>Exam Duration: <strong>45 minutes</strong></li><li># of Questions: <strong>40–60 questions</strong></li><li>Passing Score: over 700 out of 1000.</li><li>Question Formats: Multiple-choice, True/False, Drag-and-drop, Case studies</li><li>No Penalty for Guessing: There’s no negative marking, so make sure to <strong>attempt all questions</strong>.</li></ul><h4>CheatSheet</h4><p>Here’s my <a href="https://drive.google.com/file/d/1ElcyCTwHUVYVZzfGR4dOIpYwd2L4m_Ik/view?usp=sharing">cheatsheet </a>to review after practice, since I already had knowledge on ML models, AML and blob, so I just took note more on the questions I got wrong with. Hope it may help~</p><p>Good luck to anyone preparing for the AI-900 exam! If you have any questions or need further advice, feel free to leave a comment.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d35a776eb9a3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Computer Vision — Anchor Boxes如何改變目標檢測]]></title>
            <link>https://sharkyun.medium.com/computer-vision-anchor-boxes%E5%A6%82%E4%BD%95%E6%94%B9%E8%AE%8A%E7%9B%AE%E6%A8%99%E6%AA%A2%E6%B8%AC-8d6c403495d8?source=rss-2aa70868c3aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/8d6c403495d8</guid>
            <category><![CDATA[yolo]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[object-detection]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[SharkYun]]></dc:creator>
            <pubDate>Thu, 09 Jan 2025 01:34:46 GMT</pubDate>
            <atom:updated>2025-01-09T01:34:46.873Z</atom:updated>
            <content:encoded><![CDATA[<h3>Computer Vision — Anchor Boxes如何改變目標檢測</h3><h4>目標檢測在Anchor Boxes發明前後的區別Fast R-CNN vs Faster R-CNN; YOLO v1 vs v2</h4><ul><li>What is Anchor Box</li><li>Anchor Box 原理</li><li>目標檢測在Anchor Boxes發明前後</li><li>Anchor Box 範例詳細內容</li></ul><h3>什麼是Anchor Box</h3><p>Anchor box（錨框）是在深度學習中的 <strong>目標檢測（Object Detection) </strong>任務中使用的一種預定義的矩形框。它們用於幫助模型在圖像中預測物體的位置和大小，是許多現代目標檢測算法（如Faster R-CNN、SSD和YOLO v2）中的核心概念。</p><h3>Anchor Box 原理</h3><h4>Anchor Box的作用</h4><ol><li>預定義候選框： 錨框充當候選檢測框，用於覆蓋圖像中的可能物體位置。每個錨框有特定的尺寸和縱橫比。</li><li>多尺度檢測： 通過在每個位置使用多種尺寸和比例的錨框，模型可以同時檢測大小和形狀各異的物體。</li><li>加速訓練與推理： 預定義錨框減少了需要學習的參數，簡化了模型的輸出，使訓練和推理更高效。</li></ol><h4>工作原理：</h4><ol><li>設置錨框： 在特徵圖（通常是卷積神經網絡的輸出）上的每個位置，放置多個錨框，每個錨框有不同的尺寸和縱橫比。</li><li>匹配真實框： 在訓練過程中，錨框與真實的標註框（Ground Truth Boxes）進行匹配。計算錨框與真實框的 <strong>IoU（Intersection over Union）值</strong>，選擇IoU較高的錨框作為正樣本，其餘為負樣本。</li><li>迭代優化： 模型學習輸出每個錨框的調整參數（偏移量）和物體存在的概率。損失函數包括位置回歸損失和分類損失。</li><li>預測階段： 在推理時，模型對每個錨框進行調整，生成最終的預測框，並通過非極大值抑制（NMS）過濾重疊的框。</li></ol><h3>目標檢測在錨框出現前後的方法</h3><p>使用錨框的主要目的就是<strong>為了取代每次需要生成候選區域（Proposal Boxes）的步驟，從而提高檢測效率和精度</strong>。</p><p>我們需要先了解在錨框出現之前，目標檢測算法是如何處理的。這包括 Fast R-CNN 和 YOLO v1 等方法; 再來比對Faster R-CNN和YOLO v2是如何導入錨框的。</p><h3>Fast R-CNN vs Faster R-CNN</h3><h4>Fast R-CNN</h4><ul><li>Fast R-CNN 使用傳統的計算機視覺方法， <strong>Selective Search 來生成候選區域（Proposal Boxes）</strong>。</li><li>存在問題：計算效率低：計算量大，速度慢，無法滿足實時性要求。</li></ul><h4>Faster R-CNN</h4><ul><li>Faster R-CNN 引入了 Region Proposal Network (RPN)，這是一個小型的卷積網絡，能夠在特徵圖上生成候選區域。<strong>RPN 使用 Anchor Box </strong>來生成高質量的候選區域。實現端到端的訓練，提高檢測速度來達到實時性。</li></ul><h3>YOLO v1 vs YOLO v2</h3><h4>YOLO（You Only Look Once）v1</h4><ul><li>核心思想是將目標檢測轉化為單次回歸問題</li><li>將圖像劃分為 S x S 的網格（如 7x7）, <strong>每個網格負責預測固定數量（如 2 個）的邊界框</strong>，以及這些框內包含物體的置信度。</li><li>存在問題：固定的網格和邊界框數量限制了對不同尺度物體的檢測導致對多尺度物體檢測能力有限。</li></ul><h4>YOLO v2</h4><ul><li><strong>使用錨框：</strong> 在特徵圖上預設錨框，模型學習對錨框進行調整改進了對小物體的偵測能力。</li><li>高效的特徵提取： 引入了更深的網絡結構（如 Darknet-19）以獲得更豐富的特徵。</li><li>多尺度訓練： 提高了對不同尺寸的物體的檢測能力。</li></ul><h3>錨框的優勢</h3><ol><li>提升檢測精度： 通過在每個位置預測多個錨框，可以更準確地定位物體，特別是對於密集排列和小尺度的物體。</li><li>提高計算效率： 錨框方法取消了傳統的候選區域生成步驟，減少了計算量。</li><li>端到端學習： 模型可以聯合學習特徵提取、候選區域生成和最後的分類與定位，整體性能更好。</li><li>靈活性： 可以根據數據集的特性調整錨框的尺寸和比例，以適應不同的檢測任務。</li></ol><h3>Anchor Box 範例</h3><p>假設我們在4x4特徵圖的一個位置放置了三個錨框</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/492/1*e0yUC0wTld6Fwytm7ruJUQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/825/1*Sv_XtU0ACbK2RA159nlc2A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/210/1*bx0vdZ6XFp9J5q2uA7CcAg.png" /></figure><ul><li>每個Anchor Box的中心點預設位於網格單元的中心。</li><li>後續還可以透過<strong>預測偏移量</strong>，模型可以調整Anchor Box的中心點位置</li></ul><h4>對每個Grid對放下Anchor Boxes</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/603/1*u-fFlvKRz8PHGG4o_V4eJA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/675/1*D8y-6ybQQ1dLwg4mAsy-ag.png" /></figure><p>計算每個anchor box與ground truth 的IoU, 調整信心值來取得好的預測框, 但因為每個grid多數時會有重疊得anchor box, 可以使用非極大值抑制（NMS）來處理重疊的預測框，使用NMS算法，只保留置信度最高的框，移除重疊度高的其他框。</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mx6MuaHJubs-SmsQcwo57A.png" /></figure><p>從例子中我們可以看出，<strong>Anchor Box的尺寸選擇在很大程度上決定了最終輸出的好壞</strong>。所以通常不會使用手動輸入anchor box size, 大多時候是根據訓練集標籤的平均尺寸產生的。透過使用聚類演算法（如K-means）在訓練資料集上進行分析，可以確定資料集中物體的常見尺寸和比例。透過這種方式，可以選擇一組Anchor Box，使其更好地適應資料集中的物體分佈。</p><p>除了解決固定大小的問題，Anchor Box的中心點調整也非常重要。透過模型預測的偏移量，Anchor Box的中心點可以在網格單元內進行調整，以更好地匹配物體的位置。除了中心點的偏移量，模型也學習預測Anchor Box的寬度和高度的縮放因子。這些縮放因子用於調整Anchor Box的尺寸，使其更接近真實標註框。</p><p>後續的一些目標偵測演算法（如YOLO v5）引入了Anchor Free的概念，試圖透過不使用預先定義的Anchor Box來簡化模型，並改善某些情況下的偵測效能。</p><h3>總結：</h3><p>Anchor box是目標檢測模型中預設的矩形框，用於幫助模型學習如何準確地定位和識別圖像中的物體。通過合適的錨框設置，模型可以有效地處理不同尺度和形狀的物體，從而提升檢測性能。錨框的尺寸和比例需要根據數據集中的物體特性進行選擇，否則可能影響檢測效果。使用大量的錨框會增加計算量，需要在精度和效率間權衡。</p><p>Reference:</p><p><a href="https://www.youtube.com/watch?v=RTlwl2bv0Tg">C4W3L08 Anchor Boxes</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8d6c403495d8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Computer Vision — Object Detection, One-Stage vs Two-Stage]]></title>
            <link>https://sharkyun.medium.com/computer-vision-object-detection-one-stage-vs-two-stage-b05dbff88195?source=rss-2aa70868c3aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/b05dbff88195</guid>
            <category><![CDATA[yolo]]></category>
            <category><![CDATA[object-detection]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[r-cnn]]></category>
            <dc:creator><![CDATA[SharkYun]]></dc:creator>
            <pubDate>Thu, 31 Oct 2024 15:35:05 GMT</pubDate>
            <atom:updated>2024-11-10T15:05:06.226Z</atom:updated>
            <content:encoded><![CDATA[<h3>Computer Vision — Object Detection, One-Stage vs Two-Stage detectors</h3><ul><li>What is Object Detection</li><li>Difference between One-Stage and Two-Stages detectors</li><li>Applications of One-Stage and Two-Stages detectors</li></ul><h3>What is Object Detection</h3><p>In Computer vision, object detection simply means identifying and locating objects within an image. Unlike image classification, which assigns a single label to an image, object detection uses machine learning algorithms to detect multiple objects and their respective locations within an image. It typically involves two tasks:</p><ol><li><strong>Object Classification</strong>: Identifying what types of objects are present in an image (e.g., cars, people, animals).</li><li><strong>Object Localization</strong>: Determining where these objects are located within the image by drawing bounding boxes around them.</li></ol><p>Below is the overview of Object Detection’s history</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sKOlTaKwAIjrBOkW.png" /><figcaption>Zou et al. 2019. Object Detection in 20 Years: A Survey</figcaption></figure><h3>Difference between one-stage and two-stage</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OLaW7jOhIMGhNQcbPESY0Q.png" /><figcaption>By <a href="https://medium.com/u/2aa70868c3aa">SharkYun</a></figcaption></figure><p>After 2014, the object detection generally split into 2 categories: <strong>one-stage</strong> and <strong>two-stage</strong> detectors. The <strong>main difference</strong> between these two is the <strong>region</strong> <strong>proposals, </strong>where<strong> </strong>two-stage detectors do generate <strong>region proposals </strong>and one-stage detectors do not.</p><h4>Two-stage detectors</h4><ul><li><strong>Stage 1</strong>: Region proposals generation, typically using selective search or the Region Proposal Network (RPN).</li><li><strong>Stage 2</strong>: Object classification and bounding box refinement on each proposed region.</li></ul><p>Two-stage approach allows for high accuracy, as the second stage refines results, but it’s generally slower because of the two-step process. The classic two-stage methods is the <strong>R-CNN</strong> (Region-based Convolutional Neural Networks) series, including Fast R-CNN and Faster R-CNN.</p><p>Below is an example of two-stages detector: R-CNN, which it extract the region proposals from the image first.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/988/0*QvJY134B-2jEXYE3" /><figcaption>R-CNN</figcaption></figure><h4>One-stage detectors</h4><p>One-stage detectors process the input image by skipping the proposal generation and directly detecting objects in a single pass through the image. The <strong>YOLO</strong> (You Only Look Once) and <strong>SSD</strong> (Single Shot MultiBox Detector) models are prime examples of one-stage detectors.</p><ul><li><strong>Single Stage</strong>: Divides the image into a grid and simultaneously predicts bounding boxes and class scores for each region.</li></ul><p>One-stage detectors are typically faster and more suitable for real-time applications, but they may sacrifice some accuracy compared to two-stage methods.</p><p>Below is an example of one-stages detector: YOLO, which it divides an image into a grid, predicts bounding boxes with confidence scores, and uses non-maximum suppression (NMS) to finalize object detections.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*sOZJ_4sIsvNS-XEo.jpeg" /><figcaption>YOLO</figcaption></figure><h3>Application</h3><p><strong>Two-stage detectors</strong>: Known for higher accuracy, especially in complex scenes, due to their ability to refine proposals. Faster R-CNN, for instance, is often preferred in applications where precision is crucial.</p><ul><li>Use case: Medical image analysis, facial recognition, satellite imagery, and other fields where high detection accuracy is essential over speed.</li></ul><p><strong>One-stage detectors</strong>: Known for their speed, making them ideal for applications where latency is the key factor.</p><ul><li>Use case: Real-time detection like autonomous vehicles and surveillance cameras</li></ul><h3>Conclusion</h3><p>Even though two-stage detectors are more accurate due to their structure, advancements in one-stage detectors have significantly narrowed the performance gap. (As of Oct 2024, YOLO v11 just released.)</p><p>End of the day, the choice between one-stage and two-stage detectors depends more on specific application needs rather than a clear trade-off in performance.</p><p>Reference:</p><p><a href="https://www.researchgate.net/publication/353284602_Semantic_Image_Cropping">https://www.researchgate.net/publication/353284602_Semantic_Image_Cropping</a></p><ul><li><a href="https://arxiv.org/abs/1905.05055">Object Detection in 20 Years: A Survey</a></li><li><a href="https://blog.roboflow.com/what-is-r-cnn/">What is R-CNN?</a></li><li><a href="https://arxiv.org/abs/1506.02640">You Only Look Once: Unified, Real-Time Object Detection</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b05dbff88195" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>