<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>RF-DETR on </title>
    <link>https://juanmanuel.petrer.eu/en/tags/rf-detr/</link>
    <description>Recent content in RF-DETR on </description>
    <generator>Hugo</generator>
    <language>en</language>
    <lastBuildDate>Sat, 04 Jul 2026 00:00:00 +0000</lastBuildDate>
    <atom:link href="https://juanmanuel.petrer.eu/en/tags/rf-detr/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>vcalib: differentiable calibration filters for RF-DETR under illumination shift</title>
      <link>https://juanmanuel.petrer.eu/en/proyectos/vcalib/</link>
      <pubDate>Sat, 04 Jul 2026 00:00:00 +0000</pubDate>
      <guid>https://juanmanuel.petrer.eu/en/proyectos/vcalib/</guid>
      <description>&lt;blockquote&gt;&#xA;&lt;p&gt;When the lighting changes, a detector trained under other conditions degrades. The textbook answer is to collect data, label it, retrain and redeploy. vcalib explores the cheap answer: keep the model frozen and fit, in minutes and without labels, a tiny filter that corrects the input.&lt;/p&gt;&lt;/blockquote&gt;&#xA;&lt;h2 id=&#34;the-problem&#34;&gt;The problem&lt;/h2&gt;&#xA;&lt;p&gt;RF-DETR is a state-of-the-art real-time detector and, like almost any vision model, it is sensitive to illumination: dawn vs. dusk, indoor vs. overcast, a new camera, a new site. When the input distribution drifts away from the training one, detection quality drops. Retraining or fine-tuning is expensive, needs labels and is impractical on edge devices.&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
