CVAT.AI

Dataset of the Week: DUST 🛰️ Earth's orbit is becoming more congested every year, with megaconstellations like Starlink, OneWeb, Guowang, and Qianfan adding satellites alongside decades of debris, dormant satellites, and rocket bodies. Those resident space objects (RSOs) present a tremendous threat, because they move at hypervelocities and can destroy a functioning spacecraft on impact, as the 2009 Iridium-Cosmos collision showed. That's why tracking and cataloging RSOs is a central part of Space Situational Awareness (SSA). However, most existing SSA datasets are either synthetic, restricted to military and commercial use, or don't reflect real orbital conditions. DUST (Dual-Use Star Tracker), built by researchers at York University Department of Earth and Space Science, is different. It's drawn from near-infrared images captured by the Fast Auroral Imager aboard the CASSIOPE spacecraft between January and August 2023, originally an auroral science instrument that happens to also catch RSOs transiting its field of view. The dataset includes 1,378 frames and 4,237 manually verified RSO instances across 160 transits and 22 observation sessions. Twelve sessions were annotated manually in CVAT, frame by frame, exported in YOLO format. The remaining ten used a custom semi-automated pipeline with expert review. All annotations are provided in both YOLO and MOT formats. What makes this data genuinely hard: most RSOs are just a few pixels wide, smaller than the "tiny object" threshold used in benchmarks like AI-TOD. Baseline tests with OrbitTrack and an EfficientDet model confirm it's a real challenge, not just a synthetic one. If you're working on tiny object detection, multi-object tracking, or space situational awareness, this one's worth a look. 📝 Paper: https://lnkd.in/dS6mVUk2 📚 Dataset: https://lnkd.in/d4nPkNW6 Kudos to Vithurshan Suthakar, Perushan Kunalakantha, Regina Lee, and Gunho Sohn.

Microsoft is retiring Azure ML Data Labeling on September 30, 2026. If you're an #AzureML user who relies on it for annotation, it's a good time to start evaluating alternatives and planning your migration. To save you research time, we compared Azure ML and CVAT across 9 key criteria: - Dataset management & formats - Project setup & team collaboration - Annotation UX - AI-assisted annotation - Quality assurance tools - Analytics - Integrations & ecosystem fit - Deployment (cloud vs. self-hosted) - Pricing Read the full side-by-side comparison 👇 https://lnkd.in/dsuPf4Hj

If your ML pipeline runs on Azure Machine Learning Studio, this one is for you. Microsoft has officially announced the retirement of Azure ML Data Labeling on September 30, 2026. On that date, the service will be shut down, all active labeling workloads will be terminated, and all associated application data will be permanently deleted. If you're already looking for an alternative labeling solution, we put together a migration guide that covers how to move your datasets and annotations from Azure ML Studio to CVAT. It includes object detection and instance segmentation projects, Azure Blob Storage integration, annotation export from Azure ML, task creation in CVAT, and annotation import. 👇 Step-by-step migration guide: https://lnkd.in/dvEVwsfx 💬 If you have any questions about the migration, feel free to reach out to us at support@cvat.ai. We'd be happy to help you migrate.

When you show the same image to two skilled annotators, they will not always draw the same box, pick the same class, or resolve the same edge case the same way. That's not a failure. It's just how human labeling works. Left unmanaged, those differences become noise baked into your training data. Consensus is a systematic way to collect multiple opinions on the same asset, measure how much they agree, and merge them into a label you can trust. There's more to it than just assigning the same image twice, though. Where you apply it, how you set thresholds, and what you do with disagreement all matter. We wrote about how consensus works in practice, where it pays off, and the mistakes that show up again and again in real pipelines. 👇 Read the guide: https://lnkd.in/dGy73TdQ

Dataset of the Week: PMOF 🚌 When we talk about perception for autonomous vehicles, we almost always mean what's outside. But for autonomous public transport, the inside matters just as much — who's on board, are they seated safely, did someone fall? Without onboard staff, the vehicle itself needs to know. And that's a hard problem: confined spaces, shifting light, motion blur, and occlusion make in-vehicle perception genuinely difficult. To solve it, you need data that actually looks like the inside of a moving vehicle. Collected aboard the MONOCAB OWL — an autonomous rail vehicle developed in Bielefeld, Germany — by researchers from Hochschule Bielefeld and Bielefeld University, PMOF is the first publicly available overhead fisheye dataset captured under real driving conditions. A ceiling-mounted fisheye camera recorded 19,696 frames across 31 rides with 67 passengers at up to 20 km/h. The dataset supports object detection, multi-object tracking, and action recognition. Two annotators labeled all frames using CVAT, producing rotated bounding boxes in MS COCO format across three classes — person, clothing, and bag — with per-person action attributes: seated, seated on the ground, standing, and lying. The benchmark split provides 16,405 training frames and 2,940 validation frames. The authors also introduce a rotation-aware augmentation pipeline that preserves fisheye geometry, and show that combining PMOF with an office-based fisheye dataset achieves 94.8% AP50 on PMOF and 96.5% on an unseen domain. If you're building perception systems for autonomous public transport, or working on overhead fisheye detection more broadly, this dataset fills a real gap. 📝 Paper: https://lnkd.in/dsknMXMS 📚 Dataset: https://lnkd.in/d42AxG-G Kudos to Stella Wermuth, Dr. Qazi Arbab Ahmed, Klaus Neumann, and Thorsten Jungeblut.

Dense scenes — crowded streets, forest canopies, warehouse floors — are hard to navigate. When dozens or hundreds of shapes overlap in a single frame, finding the right object, selecting it, and keeping shapes in the correct order can quickly slow down the annotation workflow. In this video, we walk through the new layer controls in CVAT and show how they help you manage overlapping shapes in dense scenes. 👇 See it in action, and let us know: what's the trickiest part of annotating dense scenes for your team? #ComputerVision #DataAnnotation #CVAT #MLOps

💥 New feature alert 💥 Annotating dense scenes — crowded streets, aerial imagery, medical scans — means a lot of overlapping objects. Finding, selecting, and organizing them one by one slows down both annotation and QA. We've overhauled how layers work in CVAT to give you full control over complex annotation stacks. ↗️ The new Layer stack view in the Objects sidebar lets you see all annotations grouped by layer. Sort, drag, and drop to reorganize, or move an annotation to a specific layer directly from the action menu. ↗️ Layer numbers are now visible in the sidebar and on the canvas. You can filter annotations by layer number to declutter your workspace during QA — useful when a single frame has dozens of overlapping shapes. ↗️ You can also move layers up or down, insert new ones between existing layers without breaking the order, merge layers together, or collapse them to keep complex tasks easier to navigate. Click any annotation and CVAT will automatically open and highlight its layer in the sidebar. All layer operations, including bulk moves, are undoable. 👉 Learn more: https://lnkd.in/dZtWRSZZ ✅ Available in CVAT Online, Enterprise, and Community edition.

Most annotation guidelines look fine on paper. They list the classes, describe the objects, include a few examples. Then annotation starts, and the edge cases appear. Should a partially occluded car get a full bounding box or just the visible portion? Where exactly does the road end? Is a reflection of a vehicle still a vehicle? If the guidelines don't answer those questions, every annotator will. Differently. Based on dozens of labeling projects we've run here at CVAT, Kais Mter, our labeling services project manager, shares the most common mistakes teams make during guidelines setup and the key components of production-ready labeling guidelines. 👇 Read the guide: https://lnkd.in/eBw3Ps4F

Herbarium specimens + computer vision = a surprisingly powerful combo. 🔬 Fizza U., a Visual Analytics Intern at 🇳🇱 Naturalis Biodiversity Center, recently used CVAT to annotate botanical imagery sourced from Wikipedia and shared a great breakdown of the experience. We're always inspired to see CVAT applied to biodiversity informatics and scientific research. If you're working on niche or domain-specific annotation projects, we'd love to hear about them 👇

🔬 According to the World Health Organization, approximately 200,000 to 255,000 people die from asbestos-related diseases every year. One key line of defense is accredited lab testing, where analysts identify and count asbestos fibers in scanning electron microscope images. These fibers can be as thin as 100 nanometers and span hundreds of images per sample. Manual review at this scale is simply not sustainable. See how Andre Kempe, the CEO of ProMetronics, is building an end-to-end asbestos detection pipeline with CVAT Online, annotating ~100,000 fibers at polygon precision, training a production Vision AI model, and delivering compliance-ready lab reports automatically to ~300 accredited labs across Europe. 👇 Read the full case study https://lnkd.in/drxBTH7W #ComputerVision #DataAnnotation #IndustrialAI #MLOps #LabAutomation