CVAT.AI

Dataset of the Week: Veridis 🌽 AI is transforming agriculture, yet field monitoring remains labor-intensive due to variable conditions. To help farmers, businesses deploy mobile robots with intelligent perception to autonomously collect and analyze field data. But to do it right, those robots need relevant data and lots of it. Today's dataset is one example of such a training set. Taken in Pobladura de Fontecha, León, Spain by researchers from the Department of Mechanical, Computer Science, and Aerospace Engineering at the Universidad de León, Veridis contains 10,080 images of corn and beet fields captured by a ground mobile robot under natural daylight conditions between 9:00 AM and 1:00 PM. The dataset follows the YOLO object detection format, providing 9104 training images, 493 validation images, and 483 test images. Annotators used CVAT to mark all images with bounding boxes in normalized YOLO format and distinguished between two crop classes: beet and corn. To enhance model robustness, the creators also generated augmented versions through geometric transformations and photometric variations. They implemented privacy protection measures by automatically detecting and anonymizing people. If you work in precision agriculture or develop intelligent crop monitoring, plant health, or autonomous inspection systems, this dataset is worth trying out. 📝 Paper: https://lnkd.in/dNr6V5Zt 📚 Dataset: https://lnkd.in/ddEkdjHG Kudos to Sergio Sánchez de la Fuente, Luis Prieto López, Francisco J. Rodríguez Lera, Vicente Matellán Olivera, Miguel González-Santamarta and others.

🖌️ Manual labeling is where every great dataset starts. But as projects grow, so does the pressure on your team. Some teams take it a step further: label data in CVAT, train a custom model on it, then use that same model to pre-label the next batch. Each iteration gets faster and more accurate. So, where do you start if manual labeling is all you know? We broke down how automated labeling works, when to use it, and how to set it up in CVAT. 👇 Read the guide https://lnkd.in/dSt6JUUP

A team from Amirkabir University of Technology - Tehran Polytechnic built and released a 24K+ image obstacle detection dataset using CVAT for their original street-level annotations. Open data, solid methodology, and a second reviewer on every original capture. This is how quality labeling gets done 😎👇

🤔 How do you know your annotations are actually correct? Manual review helps, but it's slow, expensive, and hard to scale. Ground truth (GT) jobs offer a better way. You validate annotations automatically by comparing them against a small set of pre-verified benchmark labels. Errors surface faster, quality scores become objective, and reviewers focus on diagnosing root causes instead of checking boxes. Not many teams use it, and even fewer understand how much time it saves. We wrote a guide breaking down what GT labeling data is, how to build it, and how to use it for QA in CVAT. 👇 Read it here https://lnkd.in/dSR3cQan

🇨🇦 In Canada, zebra mussels cause millions of dollars in damage every year. They clog power stations and water treatment plants, damage watercraft, and devastate local ecosystems. Catching infestations early is one approach. But it requires underwater visual confirmation, and almost no public datasets exist for identifying individual, early-stage mussels in Canadian freshwater. See how Veronica Romero-Rosales, CEO of Robonotic, is turning the tide by diving into invaded Quebec lakes to collect data firsthand, labeling it with CVAT Online, and training a custom YOLO detection model. All while getting her masters in oceanography and a diving license in the process. ❤️ 👇 Read the full case study https://lnkd.in/dqfP4AJ8

💥 New feature alert 💥 Annotating dense scenes where objects share edges (buildings, land parcels, road markings) means a lot of time spent manually aligning polygons. Even small gaps or overlaps create correction work later. We've just shipped three updates that make precise polygon labeling significantly faster! ↗️ Snap to Contour (Ctrl+A) lets you align a polygon's edge to the boundary of an existing one. Place your first and last points on the boundary of a neighboring polygon, and CVAT automatically traces the correct contour between them. If it picks the wrong route, add one more point to guide it. That's all it takes. ↗️ Snap to Point (Ctrl+P) works at the vertex level. Pick a point on one polygon and it snaps automatically to the nearest point on an adjacent shape. Useful when you need to close a gap or align two polygons at a specific corner. ↗️ And once your polygons are aligned, you can now merge them too. The Join tool (J), previously available for masks only, now works for polygons as well. Select two overlapping shapes, press J, and they become one. Together, these three features cover the most common friction points in dense polygon annotation: misaligned edges, mismatched vertices, and fragmented shapes that should be a single object. ✅ Available in all CVAT editions: Community, Online, and Enterprise.

Everyone's talking about smart annotation right now. Yet nobody talks about the quality of those auto-generated labels or the QC process behind them. 🌚 Most people starting out with annotation QC think it's just... reviewing labels. But there's actually a lot more going on. Like measuring consistency across multiple annotators. Or catching errors that automated checks can't detect. If you've never built a QC process before, it can get overwhelming pretty fast. That's why we put together a guide that breaks it down into 5 layers — from label-level review to workflow controls — so the whole thing feels a lot less daunting. Worth a read if you're just getting started with annotation QC or if your current process needs a reset. 👉 https://lnkd.in/dKcqPRN7

We're thrilled to see CVAT featured in the GeoAI Toolkit for Urban Planners — a new report by UN-Habitat (United Nations Human Settlements Programme), developed through UNITAC Hamburg in partnership with ICLEI Europe. The 60-page report explores how GeoAI is reshaping urban planning across land use, mobility, waste and water management, climate resilience, public safety, and governance. It features 15 tools shaping this space, and we're glad CVAT made the list. One project that really stands out is the initiative on detecting illegal waste dumpsites along the Motagua River in Guatemala, the country's largest river, which carries significant amounts of human-made waste into the Atlantic Ocean, creating an ongoing environmental and geopolitical tension with Honduras. Co-led by UNDP Guatemala Accelerator Lab and UNDP Istanbul International Center for Private Sector in Development (UNDP ICPSD), the project combined satellite imagery, GIS, and convolutional neural networks to identify illegal dumpsites, significantly cutting the time and resources required compared to traditional monitoring. CVAT was used to label and annotate the training data that powered the project's computer vision pipeline. A big thank you to everyone behind this work — UN-Habitat (United Nations Human Settlements Programme), UNITAC Hamburg, ICLEI Europe, United Nations Office of Information and Communications Technology, HafenCity Universität Hamburg (HCU), SDG AI Lab, UNDP Accelerator Lab Guatemala, and Politecnico di Milano. And thanks to Sustainable Design Network for bringing it to our attention! 🙌

💥 New feature alert 💥 Need a quick way to pull project, task, or job metadata out of CVAT without copying it manually from the dashboard? You can now export metadata from the Projects, Tasks, and Jobs pages as CSV. That means you can take the information already visible in CVAT — such as names, assignees, statuses, dates, and related project or task details — and use it for reporting, coordination, or internal tracking outside the platform. And if you apply filters first, CVAT exports only the filtered view, so it’s easier to get exactly the list you need. A small update, but a useful one for project managers, annotation team leads, QA leads, and anyone keeping annotation workflows organized. ✅ Available in all CVAT editions: Community, Online, and Enterprise. https://lnkd.in/dhtg9pxY

Hey ML people! We come bearing great news ! 📣📣📣 We've just added support for SAM 3 assisted labeling via text prompts! This means, you can now use your existing project or task labels as textual guidance for SAM to find and label not just one object, but all recurring objects in a frame with just one click (or bbox). Yes, you heard it right — ALL RECURRING OBJECTS! 👀 This workflow works great if your datasets have many (10+) recurring objects of one or multiple classes, and they are well separated and don't overlap too much. What's cool about this feature is that your labeling workflow can now shift from object by object to class by class labeling, which not only saves time, but makes the whole experience more structured and easier to scale. The new integration is already live and available in CVAT Online and Enterprise editions. You can try it for free in demo mode in CVAT Online Free plan, and full production mode in Solo, Team and Enterprise plans. Check out the demo below, try it out, and let us know what you think! 🙌 Learn more about the new SAM 3 + text prompts integration: https://lnkd.in/dqsH89zR