Roboflow

Wanted to validate a quick idea so I prompted our new Agent and it built this for me from scratch in 10 minutes (left and right lanes are marked from driver's pov). Three months ago, if you told me I’d be able to do this so quickly, I’d politely call you a liar. A lot of production-deployed CV work has always been plumbing between pieces. We're now at a point where that layer gets easier, and teams get more room to focus on the parts that determine whether a model performs exceptionally well in production: data quality, training configs, and deployment. more coming soon on this front ... stay tuned!

Made a quick video showing part of the workflow behind a baseball computer vision project I've been working on in my free time Driveline Baseball Enterprises, Inc. I walk through how I used Roboflow as one part of the pipeline for uploading images, organizing and versioning the dataset, and then carrying that into our own training and post-processing workflow. The point of the video isn’t really the tooling by itself, it’s more about showing how the whole process connects. Going from raw image data, to dataset management, to training, to the final visual output is something that never really gets much love. I think a lot of people only ever see the finished result, but the interesting part is how all the pieces actually tie together to make that result possible. In this case, that meant turning labeled footage into a system that could eventually play a role in our broader tech stack. Cool project to be a part of. 😎 #ComputerVision #SportsTech #BaseballTech #MachineLearning #Python #DataPipeline #MLOps #AI #Baseball

You can also run Segmentation with the full Open-source RFDETR model from Roboflow in real-time 🔥🔥 A few lines of code and you have the model running and its the best mask around the objects I have seen in such a fast model even for the Nano version. Its based on the transformer architecture so you get the awesome level of detail and also very fast inference. from rfdetr import RFDETRSegMedium model = RFDETRSegMedium() detections = model.predict(frame_rgb)

A Roboflow é uma plataforma de visão computacional que permite organizar bases de imagens, anotar objetos, treinar modelos e realizar deploy de inferência via API ou edge. Quero compartilhar nosso artigo publicado no blog da Roboflow, que apresenta um case real de automação da contagem de conídios com o uso de visão computacional. No artigo, mostramos como uma atividade manual, repetitiva e sujeita a variações foi transformada em um processo digital, padronizado e muito mais escalável, com o apoio do Roboflow e do Lovable para otimizar a rotina laboratorial e tornar a tomada de decisão mais ágil. Os resultados vão além da eficiência operacional. O case destaca ganhos em padronização, rastreabilidade, redução da subjetividade e liberação das equipes para atividades de maior valor agregado. Para profissionais que atuam com qualidade, laboratório, indústria, inovação, dados ou inteligência artificial aplicada aos negócios, a leitura é bastante recomendada. https://lnkd.in/dEwYBZFk

Next week, join me and Vishrut Kaushik of Peer Robotics for a live conversation about integrating computer vision with industrial robots. We'll check out their automated movement systems, how visual intelligence unlocks additional capabilities for them, and lessons learned by Vishrut's team along the way. Register here: https://luma.com/kj3h0mwv

Could you win the World Cup with one kick? Fabrizio Guccione Guccione and I will help you find out! Thank you to Roboflow and Michael Culleton for your help with Computer Vision and TouchDesigner!

Human motion tracking systems typically follow a top-down pipeline (detect → crop → estimate). In practice, especially in lab setups or tools like #freemocap and #pose2sim, this means a person detector runs independently, and the pose model operates on fixed-resolution crops. There is a licensing problem here. Since YOLOv5, Ultralytics models use AGPL-3.0. If you build proprietary commercial software (e.g., motion tracking SaaS), you must either open-source your system or purchase an enterprise license. For that reason, many open-source pipelines (#rtmlib, #sports2d, #pose2sim) still rely on YOLOX (2021), the last Apache-2.0 YOLO variant. After ~1.5 years of using YOLOX in motion tracking setups, I’ve found it poorly suited for high-quality lab tracking. It is highly sensitive to object orientation and produces significant frame-to-frame box jitter, even for nearly static subjects. That instability propagates to pose outputs. You can smooth it offline or add causal filters online, but then you introduce lag. For real-time use cases (e.g., VR animation), that trade-off is undesirable. A better alternative is now available: RF-DETR (ICLR 2026, Apache-2.0) by Roboflow. In my side-by-side comparisons against YOLOX, it is noticeably more stable in low-motion scenes, with far less bounding box wobble. It also avoids NMS, eliminating manual tuning and associated false positives. While YOLOX can be faster, detector stability often matters more for downstream pose quality than raw FPS. To make adoption easier, I built #OpenDetect: a minimal wrapper around RF-DETR (and YOLOX) using ONNX Runtime, with CUDA, TensorRT, and Apple acceleration supported out of the box. While I focus on person class for pose estimation, detection for all COCO classes is supported. Apache-2.0 license, free for commerical use. No obligation to make your code public. 🚀 GitHub: https://lnkd.in/dFwwg4e5 🤓 Docs: https://lnkd.in/d67qCQKj If your subject is mostly static and your keypoints still jitter, improve the detector first. A stable detector can clean up your entire pose pipeline without modifying the pose model. #computervision #poseestimation #opensource