Roboflow

We are hosting the Visual Intelligence Summit on October 22 in SF. The physical world is AI's biggest opportunity and the best way to advance humanity in our lifetimes. Yet almost everyone building with AI today is building for the screen. The Summit is for the people who see it differently. If you want to be at the frontier of bringing AI into the physical world, this is your event. We built the day around two kinds of sessions. Showcase sessions put you in front of the people running AI in the real world. You'll see how their deployments work and hear from the folks who built them. Build sessions are hands-on. You'll join a small group, open your laptop, work alongside people tackling the same challenges you are, and leave having made something with the latest technology. We'll be announcing speakers soon, including builders from Tesla, Meta, Standard Bots, OpenCV, and Peer Robotics.

I used to play basketball in middle school for fun, but I hadn't played in few years. My form got rusty, and I wanted a way to rebuild my jumper without relying on guesswork. So I built a free, local shot tracker using computer vision. 🏀 Here is how it works. You mount an old phone on a tripod at the edge of the court. A custom RF-DETR model tracks the ball and rim, while zero-shot keypoint tracking monitors your wrist, elbow, shoulder, hip, and knee mechanics. The local engine calibrates distance automatically using the regulation 18-inch rim, measures release velocity, calculates arc angles, and logs makes and misses. You get an annotated video and a local web dashboard showing your form, some numbers, and joint angles at release. Everything runs locally using Roboflow's free tier, so there are no cloud API bills and your footage never leaves your machine. My favorite detail is how the engine handles net occlusions. When a shot clears the hoop, the net covers the ball for a frame or two. So instead of guessing if the ball went in or went behind the net (as an airball), the system checks how fast the ball drops when it reappears. Drag slows down a ball when it touches the net to between 10% and 90% of freefall speed, while an airball passing behind the rim stays at 100% of normal freefall. P.S. Go easy on my jumper in the video. It's been a minute since I've played 😅 Here's the whole build with all the code, the physics engine, and the local web dashboard here: 🔗 https://lnkd.in/gZNxPQiA

What changes when a detector gives you a pixel mask instead of only a bounding box? This 8-minute explainer shows how RF-DETR-Seg works, how to run it in Python, and how to read its results without overclaiming the evidence. RF-DETR and RF-DETR-Seg are developed by Roboflow. I break down the DINOv2 backbone, multi-scale projection, deformable attention, and direct set prediction. Then we walk through the article's three-step image/video workflow and inspect masks on animals, dense streets, and sports footage. Two important boundaries: - The published latency and AP numbers are scoped to the cited COCO comparisons and the article's NVIDIA T4 / TensorRT 10.4 / CUDA 12.4 / FP16 / batch-1 setup-not universal end-to-end video throughput. - The demonstrations run independent frame-by-frame inference. They do not include a temporal tracker. You'll learn: - Why masks capture shape and overlap that boxes miss - How RF-DETR performs direct set prediction without an external NMS stage - How to interpret the benchmark plots - How to install, load, and call predict() for images or video frames - Where false positives and domain validation still matter Read the full article and get the complete code: https://lnkd.in/gMXvpdKD #RFDETR #InstanceSegmentation #ComputerVision #ObjectDetection #DeepLearning

Proud to have shared the stage with Nader Khalil, Joseph Nelson, Matthew Berman, and Alex Cheema for “State of the Union: Why Local, Why Now” at the Local AI Summit during AI Engineer World’s Fair. Together, we represented the growing movement toward local, open-source, and self-custodied artificial intelligence, with perspectives from NVIDIA, Roboflow, Osmantic, Forward Future, and EXO Labs. Local AI is no longer a niche. It is becoming a fundamental part of how we build, deploy, and own the future of AI. The first panel from the summit is now live and available to watch below.

A few weeks into my role at Roboflow and I've had the chance to ship two projects I'm genuinely proud of. Here's what I've been building: 1. A 3D Soccer Offside (VAR) system that runs on any footage Pro-level offside tech can cost over $1M to install, requires synchronized camera rigs around a stadium and a team of operators in a control room. I wanted to see if the same analysis could run on a single clip from any camera, for free. The pipeline takes one video and returns an offside verdict in under 30 seconds: detect the players (RF-DETR), find the pitch lines (HSV matching), reconstruct each relevant player in 3D from a single frame (SAM3D Body), and drop them onto a metric pitch where the offside plane becomes a wall you can actually see players cross. It can run on a phone video for about $2 in GPU per match. World Cup tech, working on your kid's local game shot on an iPhone. 2. Track Class Lock, a new Roboflow Workflow block that kills label flickering If you've run object detection on video, you've seen it: an object reads "bottle" one frame, "remote" the next, "bottle" again. The detections are fine, but the flicker leaks into every counter, alert, and dashboard downstream. Track Class Lock watches each tracked object and freezes its class once the detector has voted for it confidently enough. Genuine corrections still get through. A label only switches after enough consecutive high-confidence frames, but momentary misclassifications don't. It's one block added to a standard detect-and-track workflow, turning per-frame guesses into settled identities. The energy at Roboflow is contagious! Everyone jumps in to support each other and help build great things, and it's a huge part of why I've been able to move this fast so early on. Grateful to be doing this kind of work, and there's more to come. Two very different problems, one theme I care about a lot: making computer vision that actually works in the messy real world, and making it easier for anyone to use.

Heard some of you wanted to see our booth! 👀 Day 2 of #Automate2026 was 🔥🔥🔥 Who is coming to Roboflow's happy hour??