--- title: Fine-Tuning SOTA Object Detection Models on Real-World... description: A practical walkthrough of fine-tuning three state-of-the-art object detection model families (YOLO12, YOLO26, RF-DETR) on real-world, off-distribution... canonical: https://daily.dev/posts/fine-tuning-sota-object-detection-models-on-real-world-datasets-jumlvrmgq twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Fine-Tuning SOTA Object Detection Models on Real-World Datasets | daily.dev og:description: A practical walkthrough of fine-tuning three state-of-the-art object detection model families (YOLO12, YOLO26, RF-DETR) on real-world, off-distribution... og:url: https://daily.dev/posts/fine-tuning-sota-object-detection-models-on-real-world-datasets-jumlvrmgq og:image: https://api.daily.dev/og/posts/jUmlVRmgq.png og:image:alt: Fine-Tuning SOTA Object Detection Models on Real-World Datasets og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Fine-Tuning SOTA Object Detection Models on Real-World Datasets **[JetBrains](https://daily.dev/sources/jetbrains)** · 14 min read · 0 upvotes · 0 comments ## Summary A practical walkthrough of fine-tuning three state-of-the-art object detection model families (YOLO12, YOLO26, RF-DETR) on real-world, off-distribution datasets using PyCharm. The piece first reproduces COCO val2017 baselines for six checkpoints, then shows that these same pretrained models score near zero when applied zero-shot to specialized datasets covering cable damage, bone fractures, and soda bottles, because the target classes and visual distributions differ from COCO. After fine-tuning for 10 epochs on an A100 GPU, accuracy improves dramatically for cable-damage and soda-bottles (0.91-0.97 mAP50), while bone-fracture detection remains difficult (best mAP50 0.447) due to the large domain shift from natural images to X-rays. RF-DETR Base emerges as the most consistent performer across the three domains. The post also covers setting up isolated uv environments per model family in PyCharm and downloading RF100-VL datasets via Roboflow. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Why do pretrained YOLO and RF-DETR models score close to zero mAP on custom datasets like bone fractures? Pretrained closed-vocabulary detectors like YOLO12, YOLO26, and RF-DETR physically cannot output classes absent from their COCO training head, such as bone fracture. Zero-shot mAP50-95 on bone-fracture, cable-damage, and soda-bottles datasets ranged from 0.0000 to 0.0266 across six checkpoints, even though the same models score up to 0.72 mAP50 on COCO, because their target vocabulary and visual distribution are entirely different. _Anyone deploying a detector on unfamiliar data can find fine-tuning benchmarks like these curated on daily.dev._ ### How much does fine-tuning improve YOLO and RF-DETR performance on specialized datasets like cable damage or soda bottles? After 10 epochs of fine-tuning on a single A100 GPU, mAP50 scores jumped to the 0.91-0.97 range for both cable-damage and soda-bottles datasets across YOLO12, YOLO26, and RF-DETR models. Soda-bottles transferred best because the objects visually resemble COCO categories, while cable-damage showed a bigger gap between mAP50 (up to 0.93) and mAP50-95 (topping out at 0.446) due to difficulty precisely localizing thin, elongated defects. _Developers weighing detector choices for a domain-specific task can track comparisons like this on daily.dev._ ### Which object detection model transfers best to new domains after fine-tuning, YOLO or RF-DETR? RF-DETR Base was the most consistent performer across three tested domains, winning on two out of three datasets (cable-damage and bone-fracture) and closely challenging on the third (soda-bottles). Its DETR-style architecture appeared to transfer more robustly to unfamiliar domains than the YOLO12 and YOLO26 attention- and CNN-based architectures, though even RF-DETR only reached 0.447 mAP50-95 on the hardest bone-fracture dataset. _Teams comparing detector architectures for domain transfer can follow benchmarks like this via daily.dev._ ## Similar posts on daily.dev - [Best Object Detection Models for Machine Learning in 2026](https://daily.dev/posts/best-object-detection-models-for-machine-learning-in-2026-jzubdcufk) · JetBrains · 0 upvotes · 0 comments - [Fine-tune Ultralytics YOLO26 Object Detection Model](https://daily.dev/posts/fine-tune-ultralytics-yolo26-object-detection-model-ktflc0ns2) · Daily Dose of Data Science \| Avi Chawla \| Substack · 0 upvotes · 0 comments - [Top Object Detection Models for Your Projects in 2025](https://daily.dev/posts/top-object-detection-models-for-your-projects-in-2025-ei6lsr52c) · DigitalOcean Community · 1 upvotes · 0 comments --- Tags: [#devtools](https://daily.dev/tags/devtools), [#deep-learning](https://daily.dev/tags/deep-learning), [#computer-vision](https://daily.dev/tags/computer-vision), [#object-detection](https://daily.dev/tags/object-detection) [View this post on daily.dev](https://daily.dev/posts/fine-tuning-sota-object-detection-models-on-real-world-datasets-jumlvrmgq) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Fine-Tuning SOTA Object Detection Models on Real-World Datasets","url":"https://daily.dev/posts/fine-tuning-sota-object-detection-models-on-real-world-datasets-jumlvrmgq","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/fine-tuning-sota-object-detection-models-on-real-world-datasets-jumlvrmgq"},"datePublished":"2026-08-31T15:20:25.174Z","dateModified":"2026-08-31T15:20:55.679Z","description":"A practical walkthrough of fine-tuning three state-of-the-art object detection model families (YOLO12, YOLO26, RF-DETR) on real-world, off-distribution...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8b62b0bc9f5ae160c02e13d8913944b5?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8b62b0bc9f5ae160c02e13d8913944b5?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"JetBrains","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"JetBrains","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/53ecf0c676f34b3896ee109609d91efa","url":"https://daily.dev/sources/jetbrains"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/fine-tuning-sota-object-detection-models-on-real-world-datasets-jumlvrmgq","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"devtools,deep-learning,computer-vision,object-detection","timeRequired":"PT14M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"JetBrains","item":"https://daily.dev/sources/jetbrains"},{"@type":"ListItem","position":3,"name":"Fine-Tuning SOTA Object Detection Models on Real-World Datasets"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/fine-tuning-sota-object-detection-models-on-real-world-datasets-jumlvrmgq#faq","mainEntity":[{"@type":"Question","name":"Why do pretrained YOLO and RF-DETR models score close to zero mAP on custom datasets like bone fractures?","acceptedAnswer":{"@type":"Answer","text":"Pretrained closed-vocabulary detectors like YOLO12, YOLO26, and RF-DETR physically cannot output classes absent from their COCO training head, such as bone fracture. Zero-shot mAP50-95 on bone-fracture, cable-damage, and soda-bottles datasets ranged from 0.0000 to 0.0266 across six checkpoints, even though the same models score up to 0.72 mAP50 on COCO, because their target vocabulary and visual distribution are entirely different. Anyone deploying a detector on unfamiliar data can find fine-tuning benchmarks like these curated on daily.dev."}},{"@type":"Question","name":"How much does fine-tuning improve YOLO and RF-DETR performance on specialized datasets like cable damage or soda bottles?","acceptedAnswer":{"@type":"Answer","text":"After 10 epochs of fine-tuning on a single A100 GPU, mAP50 scores jumped to the 0.91-0.97 range for both cable-damage and soda-bottles datasets across YOLO12, YOLO26, and RF-DETR models. Soda-bottles transferred best because the objects visually resemble COCO categories, while cable-damage showed a bigger gap between mAP50 (up to 0.93) and mAP50-95 (topping out at 0.446) due to difficulty precisely localizing thin, elongated defects. Developers weighing detector choices for a domain-specific task can track comparisons like this on daily.dev."}},{"@type":"Question","name":"Which object detection model transfers best to new domains after fine-tuning, YOLO or RF-DETR?","acceptedAnswer":{"@type":"Answer","text":"RF-DETR Base was the most consistent performer across three tested domains, winning on two out of three datasets (cable-damage and bone-fracture) and closely challenging on the third (soda-bottles). Its DETR-style architecture appeared to transfer more robustly to unfamiliar domains than the YOLO12 and YOLO26 attention- and CNN-based architectures, though even RF-DETR only reached 0.447 mAP50-95 on the hardest bone-fracture dataset. Teams comparing detector architectures for domain transfer can follow benchmarks like this via daily.dev."}}]} ```