• 7 min read
EfficientDet vs YOLO for Object Detection
EfficientDet vs YOLO compared on architecture, speed, accuracy, and licensing, plus why RF-DETR is the stronger detector to deploy today.
• 5 min read
Launch: Process data directly from Wasabi, Backblaze B2, Cloudflare R2, and more
We’ve expanded Batch Processing to support the wider ecosystem of S3-compatible storage. Learn how to run inference on terabytes of data stored in Wasabi, Backblaze, and Cloudflare R2, and more.
• 9 min read
How to Train an EfficientNet Model with a Custom Dataset
Learn how to train an EfficientNet model with a custom dataset.
• 12 min read
What Is OpenPose? A Guide to Pose Estimation
In this guide, we discuss what OpenPose is, what you can do with the model, and how to use OpenPose.
• 13 min read
What Is YOLOv5? Architecture, Sizes, and How It Works
In this article, we discuss what is new in YOLOv5, how the model compares to YOLOv4, and the architecture of the new v5 model.
• 7 min read
How to Use the Segment Anything Model (SAM)
Segment Anything (SAM) is a computer vision model developed by Meta AI. In this guide, you will learn how to use SAM on your own data.
• 6 min read
How to Use Qwen3-VL in Roboflow
In this guide, build an image-understanding workflow in Roboflow that turns Qwen3-VL into a production-ready vision agent you can run via API or on your own GPU.
• 4 min read
AI for Live Video: Introducing the Serverless Streaming API
SUMMARY
The Roboflow Serverless Video Streaming API lets you run computer vision workflows on live video from webcams, RTSP feeds, or video files by streaming directly to the cloud over WebRTC, with no inference environment to configure or manage. Compute instances are ephemeral and scale automatically, so you pay only
• 4 min read
Introducing Roboflow Rapid: Text prompt to vision model in minutes
SUMMARY
Roboflow Rapid is a prompt-based model creation engine that takes you from a text description and a handful of images or a short video to a deployed custom model without manual labeling. You type a prompt describing what to detect, Rapid automatically finds and annotates matching objects, and
• 6 min read
Launch: Use Segment Anything 3 (SAM 3) with Roboflow
SUMMARY
Roboflow has integrated SAM 3, Meta's foundation model for Promptable Concept Segmentation, across every major part of its platform, covering the Playground for interactive testing, the inference API, edge and private-cloud deployment, Roboflow Workflows for building production pipelines, and the Label Assist annotation tool for accelerating
• 13 min read
SAM 3: Segment Anything with Concepts
Learn what SAM 3 is, what you can do with the model, see examples of SAM 3 in use, and learn how you can use the model.
• 9 min read
Best iOS Object Detection Models
Explore the best iOS object detection models: RF-DETR, YOLO11, and MobileNet SSD. Learn how to train, quantize, and deploy vision AI on iOS.
• 9 min read
How to Create a YOLOv11 Android App
SUMMARY
This tutorial walks through training a custom YOLOv11 object detection model with Roboflow and deploying it directly on an Android device to build a real-time coin counting app. The model weights are converted from a .pt file to TorchScript format so they can run on-device without a
• 4 min read
Launch: Verified Model Metrics
SUMMARY
Model evaluation scores are not always comparable across providers because teams use different test sets, confidence thresholds, and calculation methods, which makes selecting models for production unreliable. Roboflow Verified Model Metrics addresses this by calculating mAP using open-source COCO Evaluation methodology from the Supervision library on independent test
• 9 min read
How to Build an iOS App with Visual AI Capabilities
SUMMARY
This tutorial walks through building an iOS app that detects glasses in real time using a custom RF-DETR Nano model trained in Roboflow and deployed via the roboflow-swift SDK. RF-DETR Nano is chosen specifically for its low latency on live camera streams, making it well suited
• 6 min read
How to Monitor Red Zones with Computer Vision
Learn how to monitor red zones for safety compliance using computer vision.
• 11 min read
How to Train RF-DETR on a Custom Dataset
Learn how to train an RF-DETR model on a custom dataset.
• 5 min read
How to Visually Compare Computer Vision Models
Learn how to compare two computer vision models visually using Roboflow Workflows.
• 3 min read
How to Deploy Computer Vision Models to a Raspberry Pi
Learn how to deploy computer vision models to a Raspberry Pi.
• 7 min read
Deploying Machine Learning Models with PyTorch, gRPC, and asyncio
SUMMARY
Serving a machine learning model with gRPC and asyncio gives you a faster, more scalable alternative to HTTP/1 REST endpoints: messages are binary Protobuf rather than JSON, type-safe client and server stubs are generated from a .proto file, and multiple images can be batched into a single
• 4 min read
Launch: Dedicated Deployments
SUMMARY
Roboflow Dedicated Deployments provision exclusive CPU or GPU servers pre-configured with Roboflow Inference, removing the need to manage hardware, drivers, TLS certificates, or scaling yourself. Two environment types are available: a short-lived development tier billed by the hour for testing and prototyping, and a persistent production tier
• 5 min read
Count Objects on a Conveyor Belt Using Computer Vision
SUMMARY
This tutorial builds an end-to-end system for counting bolts and nuts on a moving conveyor belt, covering every stage from data collection to local deployment. Video frames are annotated using Roboflow's auto-labeling with Grounding DINO, a custom detection model is trained in the cloud,
• 10 min read
How to Use Your GPU in a Docker Container
Learn how to use your GPU in a Docker container and measure upir application's performance.
• 7 min read
What is ByteTrack? A Deep Dive.
SUMMARY
ByteTrack is a multi-object tracking algorithm that improves on earlier methods by using all detections, not just high-confidence ones, to maintain consistent object identities across video frames. Its two-stage data association module first matches high-confidence detections to existing tracklets, then reconciles lower-confidence ones using
Stay Connected