Blog

Joseph Nelson

Roboflow cofounder and CEO. On a mission to transform every industry by democratizing computer vision. Previously founded and sold a machine learning company.

Latest Posts by Joseph Nelson

teams

17 Jan 2021 • 2 min read

Teams for All: Create Free Teams on Roboflow

SUMMARY Roboflow now offers free teams, letting groups of three or fewer collaborate on datasets, annotation, and model deployment within a single shared workspace using individual API keys per account. Team members can share any dataset they own, annotate images together, and manage uploads and inference without sharing credentials. Larger

transfer

13 Jan 2021 • 4 min read

A Primer on Transfer Learning

SUMMARY Transfer learning lets you apply weights a model learned on one problem to a related problem, cutting training time and reducing the labeled data you need to get started. In practice this means initializing from a model already trained on a large dataset like COCO or ImageNet, then fine-

seed

12 Jan 2021 • 2 min read

Announcing Roboflow’s Seed Round

Computer vision is a generational technology. Like the PC, internet, and mobile phones, computer vision’s impact will reshape every industry. In transportation, for example, the advent of machine vision gives rise to self-driving cars, rewiring how we navigate the world around us. Manufacturing companies are finding and reducing

6 Dec 2020 • 2 min read

Share Your Datasets with the Computer Vision Community

SUMMARY Roboflow offers to publicly host community-contributed computer vision datasets, making them discoverable to tens of thousands of developers and automatically indexed by Google Dataset Search. Hosted datasets are available for download in every major annotation format, including Pascal VOC XML, COCO JSON, and TFRecords, with no conversion work

Scaled-YOLOv4 is Now the Best Model for Object Detection

3 Dec 2020 • 8 min read

Scaled-YOLOv4 is Now the Best Model for Object Detection

SUMMARY Scaled-YOLOv4 set a new state-of-the-art result for object detection on the Microsoft COCO benchmark at the time of its publication, surpassing Google's EfficientDet across both the speed and accuracy ends of the performance frontier. The model achieves this by scaling the depth, stages,

8 Nov 2020 • 3 min read

Hands on with the Roboflow Infer Web Application Interface Builder

SUMMARY After training a model with Roboflow Train, the Example Web App gives an immediate browser-based interface to test predictions: upload an image or provide a URL, adjust confidence thresholds and bounding box overlap, filter classes, and view results as an annotated image or raw JSON. Because the app

Evaluating Object Detection Models with mAP by Class

3 Nov 2020 • 3 min read

Evaluating Object Detection Models with mAP by Class

SUMMARY Overall mAP scores can mask weak spots in a multi-class object detection model: a 98% aggregate hides the fact that one class is dragging performance down. Roboflow Train addresses this by surfacing per-class mAP on both the validation and training sets after every training run, so you

protect endangered species

30 Oct 2020 • 5 min read

How This Fulbright Scholar is Using Computer Vision to Protect Endangered Species

SUMMARY Kasim Rafiq, a Fulbright Scholar and National Geographic Explorer at UC Santa Cruz, is building a vehicle-mounted computer vision system that automatically detects and photographs wildlife during safari game drives, then uses those detections for species population surveys at a fraction of the cost of traditional field methods.

Train Test Split Guide and Overview

28 Oct 2020 • 2 min read

Train Test Split Guide and Overview

SUMMARY Splitting a dataset into training, validation, and test sets is a core practice for detecting overfitting and ensuring a computer vision model generalizes beyond its training examples. Roboflow defaults to a 70/20/10 split at upload time but gives you direct control: you can assign individual image batches

tesla

11 Oct 2020 • 6 min read

How Tesla Teaches Cars to Stop

SUMMARY Tesla Senior Director of AI Andrej Karpathy's CVPR 2020 talk reveals that even a seemingly straightforward computer vision task like stop sign detection involves dozens of edge cases: occluded signs, non-standard mounting, signs that do not require stopping, and conditionally active signs. Tesla handles this by

Introducing Grayscale and Hue/Saturation Augmentations

27 Sep 2020 • 3 min read

Introducing Grayscale and Hue/Saturation Augmentations

SUMMARY Roboflow has added three color-based augmentation options, grayscale, hue, and saturation, to help models generalize to objects that appear under different lighting conditions or in varied color environments. Grayscale augmentation randomly converts a configurable percentage of training images to single-channel output, reducing the model's reliance

data science

12 Sep 2020 • 1 min read

Elisha Odemakinde Hosts Roboflow for a Fireside Chat

Recently, Roboflow machine learning engineer Jacob Solawetz sat down with Elisha Odemakinde, an ML researcher and Community Manager at Data Science Nigeria, for a Fireside chat. During the conversation, Jacob walks through his journey from his career in financial trading to artificial intelligence, a move from natural language processing to

Explore Images, Annotations, and Metadata Easier

2 Sep 2020 • 2 min read

Explore Images, Annotations, and Metadata Easier

SUMMARY Roboflow's updated image preview interface lets users view annotations full-screen, toggle individual annotation groups on and off by hovering, and surface image metadata including the original filename, camera used, modification date, and which split the image belongs to. For dataset versions with augmentations applied, the preview

Y Combinator

25 Aug 2020 • 2 min read

Roboflow Joins Y Combinator's Summer 20 Batch

Roboflow is enabling any developer to use computer vision (without being a machine learning expert). Computer vision is the first technology that fundamentally allows us to rewrite human-computer interaction. Until now, when using computers to interact with the world, humans would assess their environment, provide commands to a computer,

16 Aug 2020 • 4 min read

Using Computer Vision to Find Brands in YouTube Videos

SUMMARY Software engineer Charles Herring built a browser-based computer vision model that detects power tool brands appearing in DIY YouTube videos, giving companies a way to measure unpaid exposure and identify sponsorship opportunities. Roboflow replaced Google Cloud AutoML as his data management layer, providing dataset-level class distribution statistics

pp-yolo

3 Aug 2020 • 8 min read

PP-YOLO Surpasses YOLOv4 - State of the Art Object Detection Techniques

Baidu publishes PP-YOLO and pushes the state of the art in object detection research by building on top of YOLOv3, the PaddlePaddle deep learning framework, and cutting edge computer vision research.

Improving Infrastructure Asset Management with Computer Vision

21 Jul 2020 • 6 min read

Improving Infrastructure Asset Management with Computer Vision

SUMMARY Result! Data built a mobile app called Spobber that uses computer vision to detect hectometre signs along Dutch highways, verifying whether field assets match database records and flagging missing or damaged signs automatically. This guest post walks through their full pipeline: data collection from a moving vehicle, annotation, augmentation

Retail Store Item Detection using YOLOv5

18 Jul 2020 • 9 min read

Retail Store Item Detection using YOLOv5

SUMMARY This tutorial trains a YOLOv5 model on the SKU110k dataset to detect and localize products on retail store shelves, with dataset preparation handled through Roboflow and model training run in Google Colab. The resulting detector draws bounding boxes around shelf items and can support inventory tracking or automated checkout

computer vision software for human rights

1 Jul 2020 • 3 min read

How AI Acts as a Human Rights Watchdog for the Maasai People

SUMMARY A team of students built a computer vision system that processes satellite imagery from Google Earth Engine to monitor Maasai villages in Kenya and Tanzania, tracking settlement size, health, and evidence of burning to document land displacement by government and conservation interests. The model was trained using Roboflow'

Improving Uno with Computer Vision (Plus the Dataset so You Can Too)

28 Jun 2020 • 6 min read

Improving Uno with Computer Vision (Plus the Dataset so You Can Too)

SUMMARY Software engineer Adam Crawshaw built an Android app that uses object detection to automatically score Uno hands, handling the multi-round point system most players overlook. Starting from 41 source card images, he used OpenCV to synthetically composite cards into stacked groups, producing a dataset of 8,992 images

boggle

25 Jun 2020 • 7 min read

Creating BoardBoss: A Mobile Application that Improves Boggle

SUMMARY BoardBoss is a free iOS app, built by the Roboflow team, that scans a Boggle board with a smartphone camera, identifies letters in any orientation across more than 30 classes, and overlays all valid word solutions in augmented reality. The post walks through the full pipeline: collecting 355 images

How to Convert Annotations from PASCAL VOC to YOLO Darknet

19 Jun 2020 • 5 min read

How to Convert Annotations from PASCAL VOC to YOLO Darknet

SUMMARY PASCAL VOC stores bounding box annotations as XML files with absolute pixel coordinates, while YOLO Darknet uses plain text files with class index and normalized center coordinates. Converting between them means restructuring both the file format and the coordinate representation. This post walks through both formats with concrete examples

Responding to the Controversy about YOLOv5

12 Jun 2020 • 16 min read

Responding to the Controversy about YOLOv5

SUMMARY When Roboflow published an early YOLOv5 benchmark in June 2020, the machine learning community raised two substantive objections: whether naming Glenn Jocher's repository YOLOv5 was appropriate given no paper was published, and whether the benchmark methodology was reproducible and fair. This post documents the full history of

YOLOv5 is Here: State-of-the-Art Object Detection at 140 FPS

10 Jun 2020 • 4 min read

YOLOv5 is Here: State-of-the-Art Object Detection at 140 FPS

SUMMARY Released in June 2020, YOLOv5 is a PyTorch-based object detection model from Ultralytics that differs from YOLOv4 in several practical ways: it uses a lighter Ultralytics PyTorch framework instead of Darknet, its smallest variant (YOLOv5s) weighs 27 MB compared to 244 MB for YOLOv4, and it offers faster

Stay Connected

Get the Latest in Computer Vision First