• 6 min read
Zero-Shot Content Moderation with OpenAI's New CLIP Model
Learn how to use the CLIP zero-shot model to moderate visual content.
• 4 min read
Andrew Ng: "Deploying to production means you're halfway there."
SUMMARY
Andrew Ng's core argument at the Scale Transform conference is that deploying a model to production is the midpoint, not the finish line: real-world performance requires a continuous feedback loop where data collected from deployment improves the next model version. He makes the case for a
• 1 min read
Webinar: How to Build a Detectron2 Model with Roboflow and Paperspace Gradient
SUMMARY
This post announces and links the recording for a joint Roboflow and Paperspace webinar that walks through training a Detectron2 object detection model on a blood cell count dataset. The session covers accessing public image datasets through Roboflow, applying preprocessing and augmentation, and running GPU-backed training notebooks on
• 2 min read
Video Inference with Roboflow
SUMMARY
Running a Roboflow model on video works by splitting the footage into individual frames, sending each frame through the model for predictions, then stitching the annotated frames back into a video. Roboflow provides an open source shell script that handles all three steps, so you can visualize detection output
• 4 min read
Danger Monitoring for Cyclists using Raspberry Pi and Object Detection
SUMMARY
This project builds a low-cost cyclist danger monitor using a Raspberry Pi 4, a Google Coral TPU stick for faster inference, and an LED traffic light that warns the rider when cars or bicycles are detected in proximity. The model was trained on three datasets totaling over 6,
• 4 min read
What Is Embedded Machine Learning?
SUMMARY
Embedded machine learning is the practice of running inference directly on microcontrollers and low-power hardware, such as a Raspberry Pi, NVIDIA Jetson, or Luxonis OAK, rather than sending data to a cloud server. Shrinking model sizes and cheaper, faster hardware have made this practical: inference requires far fewer
• 3 min read
Using Computer Vision to Detect Personal Protective Equipment
SUMMARY
Construction site PPE compliance can be automated with an object detection model that flags safety violations, such as missing hard hats, directly from site images. Data scientist Jaco Lau built this using Roboflow for image annotation, class balancing, and model training, then deployed the detector inside a Streamlit application
• 9 min read
How We Built Paint.wtf, an AI Game with 150,000+ Submissions that Judges Your Art
SUMMARY
Paint.wtf is an online drawing game built by Roboflow and Booste that uses OpenAI's CLIP model to score user submissions against absurdist text prompts, ranking drawings by how closely their image embeddings match the prompt's text embedding. GPT-2 generated the bulk of the
• 3 min read
Introducing Roboflow Support for NVIDIA Jetson
SUMMARY
Roboflow Inference supports the full NVIDIA Jetson family, letting teams deploy custom-trained computer vision models to edge devices using three Docker commands, with GPU acceleration enabled by default and speeds up to 15 frames per second on the Jetson Xavier NX. The on-device inference server maintains API
• 3 min read
3 Things You Should Know From This Year's 2021 Startup Summit
Last week, I attended the 2021 Startup Summit from my home office in Des Moines, Iowa.
Perhaps one of the few good things to come out of this pandemic are virtual conferences, which have significantly increased the accessibility of thought leadership in our respective industries. Read about some of the
• 1 min read
Roboflow Changelog: March 2021
The monthly changelog showcases improvements in Roboflow over the past month. You can find the previous changelog here.
In February, we had a major focus on reinforcing our foundation; after several months of rapid feature expansion we spent a considerable amount of time fixing bugs, improving infrastructure, and preparing for
• 3 min read
5 Emerging Trends in Computer Vision Applications from OpenCV's AI Competition
SUMMARY
Reviewing over 200 project proposals submitted to the 2021 OpenCV AI Competition, five application patterns stand out across the more than 1,200 competing teams: assisting the visually impaired with spatial awareness, improving workplace safety by identifying hazardous zones, monitoring COVID protocol compliance, and tracking animal health and conservation
• 8 min read
Running Tensorflow JS on a NVIDIA Jetson
SUMMARY
Running TensorFlow.js on an NVIDIA Jetson requires extra steps because Google does not ship prebuilt arm64 binaries for the Node.js backend, but two practical paths exist: use the WebGL backend in Chromium for zero-configuration GPU acceleration at around 2 to 4 frames per second, or compile
• 8 min read
How to Use Roboflow and Streamlit to Visualize Object Detection Output
SUMMARY
This tutorial builds a blood cell detection app that trains an object detection model in Roboflow to identify platelets, red blood cells, and white blood cells, then wraps the model's API in a Streamlit interface so users can upload images, adjust confidence thresholds with a slider, and
• 3 min read
How to Use Google Earth Engine and Python API to Export Images to Roboflow
SUMMARY
Google Earth Engine's Python API can pull multispectral imagery from the Sentinel-2 satellite and export it as image files ready to upload directly into Roboflow for training computer vision models on geographic data. The tutorial covers authenticating to both Google Earth Engine and Google Drive, selecting
• 3 min read
Introducing Upload Batches, the Unannotated Queue, and Mark as Null
SUMMARY
Roboflow introduced three annotation workflow features to help teams manage iterative dataset collection: Upload Batches automatically groups each new image upload so teams can divide labeling work without merging separate datasets; the Unannotated Queue makes it easy to track exactly which images still need labels; and Mark as Null
• 3 min read
Using Computer Vision to Improve Disability Access in Music Education
SUMMARY
A high school student built the Mehek Box, a tactile rhythm education tool for students with disabilities, and extended it with a computer vision feature that recognizes physical rhythm blocks via an in-app camera. She trained an EfficientDet-D0 object detection model on a custom dataset built from
• 5 min read
Liquid Neural Networks in Computer Vision
Excitement is building in the artificial intelligence community around MIT's recent release of liquid neural networks. The breakthroughs that Hasani and team have made are incredible. In this post, we will discuss the new liquid neural networks and what they might mean for the vision field.
• 4 min read
Using Computer Vision to Help Win $1 Million in Mountain Dew's Big Game Contest
SUMMARY
Mountain Dew's Super Bowl LV ad challenged viewers to count the exact number of bottles in the commercial for a $1 million prize, and this post builds an object detection model to assist with that count. The 30-second video was sampled at three frames per second
• 7 min read
How to Train and Deploy a License Plate Detector to the Luxonis OAK
In this post, we will leverage Roboflow and the Luxonis OAK to train and deploy a custom license plate model to your OAK device.
• 4 min read
Using Computer Vision to Make Card Grading Faster and Cheaper
SUMMARY
Professional card grading is slow and expensive, and a large portion of the assessment (edge wear, corner wear, centering, surface damage) maps well to computer vision. This guest post describes a prototype that uses OpenCV for edge detection and card alignment checks, then trains a YOLOv5 model on a
• 2 min read
Roboflow Changelog: February 2021
Welcome to our monthly roundup post of new features and enhancements. You can find the previous changelog here.
The biggest new features released this month were transfer learning and Label Assist.
Roboflow Organize
* Announced teams for all (now teams of up to 3 people can collaborate on the Starter Plan
• 3 min read
Using Computer Vision to Win at Duck Hunt
Can we use object detection to automate identifying moving objects on a screen? Abhinav Mandava leverages Roboflow to create an aimbot (which automates aiming and firing for the player) for Duck Hunt.
• 11 min read
Using Unity Perception to train an object detection model with synthetically generated images
SUMMARY
Unity's Perception package can generate perfectly annotated synthetic images using only software, giving computer vision projects a way to bootstrap training data before real-world images are available. This tutorial walks through setting up Unity 2019.4 LTS with the Perception package, generating a labeled synthetic dataset,
Stay Connected