Ultralytics

New tutorial | Real-time inference with Ultralytics YOLO11 + Streamlit ⚡ Real-time inference is essential for many computer vision applications. In this tutorial, we walk through how to build and run a live inference application using YOLO11 and Streamlit. What'll be covered: ✅ Running the application using both CLI and Python ✅ Configuring application settings and switching models ✅ Testing real-time object detection and instance segmentation A practical guide to building interactive, real-time computer vision applications with YOLO11. Watch now ➡️ https://lnkd.in/e_CHMAXP

Train Ultralytics YOLO11 using the KITTI dataset! 🚗 The KITTI dataset remains one of the most influential benchmarks in computer vision, powering research in autonomous driving, traffic analysis, and urban scene understanding. With Ultralytics, you can load the entire dataset instantly using the ready-to-use KITTI YAML file and build models capable of understanding complex driving environments, from traffic flow to vulnerable road-user detection. Start now ➡️ https://bit.ly/485lB3t

New release Ultralytics v8.3.240 | More reliable SAM 2/3 video segmentation on Apple Silicon 🚀 This stability release improves SAM 3 on MPS, bounds long‑video memory growth, and smooths export + docs for faster day‑to‑day workflows. Minor updates: ✅ SAM 3 now runs reliably on Apple Silicon (MPS) with an MPS-safe rotary encoding path ✅ SAM 2/3 video inference memory is now bounded to reduce VRAM growth on long runs ✅ `pip install ultralytics[export]` now pulls `onnxslim` automatically for cleaner ONNX exports Ultralytics v8.3.240 release notes ➡️ https://lnkd.in/e_Qn-2dT

Crack segmentation with Ultralytics YOLO11! With the segmentation task, you can go beyond detecting surface cracks to outline their shape and length precisely, even when cracks are thin or partially hidden in noisy textures. This level of pixel-accurate detail is essential for structural health monitoring and predictive maintenance. You can also integrate this into real-time pipelines for bridges, buildings, machinery parts, and more. Read more ➡️ https://bit.ly/49t29yK

New release v8.3.239 | SAM 3 text prompts now plug-and-play 🧩 This update removes extra SAM 3 tokenizer setup, adds real-time disk/network throughput monitoring, and improves ONNX export compatibility (including Jetson JetPack 6) for smoother deploys. Minor updates: ✅ SAM 3 no longer needs manual BPE vocab downloads or `bpe_path`, aligning with the simplified API in the SAM 3 docs https://lnkd.in/eFgH7Eff ✅ `SystemLogger.get_metrics(rates=True)` adds live disk + network MB/s for better bottleneck diagnosis ✅ ONNX export now supports newer ONNX versions up to `<2.0.0`, improving Jetson/JetPack 6 workflows Ultralytics v8.3.239 release notes ➡️ https://lnkd.in/eaXRe97i

FlipUD & FlipLR: Simple augmentations making Ultralytics YOLO models stronger! 🔁 In computer vision, small transformations can create big improvements. Two of the most effective, yet underrated, data augmentation techniques are flipud (up–down flip) and fliplr (left–right flip). Why they matter: ✅ fliplr: Horizontal flip helps your model generalise when objects appear mirrored, like vehicles approaching from different directions or people facing left vs. right. ✅ flipud: Vertical flip is useful for aerial, industrial, or top-down vision tasks where objects can appear upside-down, boosting robustness in drones, robotics, and surveillance. By flipping existing images, you instantly create new, realistic variations, reducing overfitting and improving real-world performance without collecting extra data. Learn more ➡️ https://bit.ly/3JY0zur

New release Ultralytics v8.3.238 | SAM 3 concept/video segmentation is more robust and smoother to use 🧠🧩 This refinement-focused update improves SAM 3 batching stability, macOS (MPS) reliability, and installs, while making exports (TFLite/ONNX) more predictable for deployment. Minor updates: ✅ Fixed SAM 3 “encode once, query many times” prompt-batching cache bug for stable multi-prompt workflows ✅ Added SAM 3 tokenizer dependency auto-checks + improved Apple Silicon (MPS) attention reliability ✅ Hardened export paths (TFLite dtype fixes, safer ONNX FP16 guards, clearer version checks) Ultralytics v8.3.238 release notes ➡️ https://lnkd.in/ekNi5GPW

New release v8.3.237 | SAM 3 image + video segmentation now supported 🚀 Ultralytics adds full SAM 3 support for concept segmentation (text + exemplar prompts) and video tracking, plus smoother exports and day-to-day training/validation improvements. See the new SAM 3 docs https://lnkd.in/eFgH7Eff for quick start. Minor updates: ✅ ONNX export: FP16 supported on CPU with graceful fallback ✅ More resume-training overrides (e.g., `workers`, `cache`, `patience`, `val`, `plots`) ✅ More robust Edge TPU/IMX dependency handling during export Ultralytics v8.3.237 release notes ➡️ https://lnkd.in/eRgGJ_k3

New tutorial | Train Ultralytics YOLO11 on the VisDrone dataset 🚁 The VisDrone dataset is one of the leading benchmarks for drone-based object detection and aerial scene understanding. In this tutorial, we show how to train YOLO11 on VisDrone and evaluate its performance on real aerial imagery. We cover: ✅ Dataset structure, classes & YAML configuration ✅ Setting up Ultralytics in Google Colab ✅ Training YOLO11 step by step on the VisDrone dataset ✅ Running predictions on drone imagery (cars, pedestrians, bicycles & more) A complete walkthrough for building aerial object detection models with YOLO11. Watch now ➡️ https://lnkd.in/dUpuywZN