Roboflow Blog

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

18 Jun 2020 • 4 min read

How to Build a Custom Open Images Dataset for Object Detection

SUMMARY Google's Open Images dataset contains millions of labeled images across hundreds of object classes, all under a CC BY 4.0 license, but downloading the full dataset requires over 500 GB of storage. This post shows how to use the OIDv4_ToolKit to query specific classes and

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

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

3 Jun 2020 • 3 min read

Teaching a Drone to Fly on Auto Pilot with Roboflow

SUMMARY Victor Antony, a data scientist at the University of Rochester, built a computer vision autopilot system to detect gates that a drone must fly through, starting from only 300 images captured in a warehouse. Using Roboflow, he applied augmentations for rotation, lighting, and blur to expand the dataset to

2 Jun 2020 • 7 min read

Getting Started with Data Augmentation in Computer Vision

SUMMARY Data augmentation artificially expands training sets by applying flips, crops, rotations, noise, and occlusion techniques, and it has a measurable impact on object detection accuracy even when collecting more labeled images is not practical. RF-DETR relies heavily on augmentation to achieve its COCO benchmark results, and the same

31 May 2020 • 1 min read

A New Video Tutorial: YOLOv4 in PyTorch

SUMMARY This post introduces a 20-minute video tutorial covering how to use Roboflow to prepare a custom dataset and train a YOLOv4 model in PyTorch. It accompanies the team's earlier written guides on the YOLOv4 architecture and Darknet-based training, extending them to the PyTorch framework for

27 May 2020 • 3 min read

Introducing An Even Better Way to Preview Image Preprocessing and Augmentation

SUMMARY Roboflow's revamped preprocessing and augmentation previews let you see exactly what each step does to your images before committing to a pipeline, replacing guesswork and boilerplate transformation code with a visual interface. The update covers choices like aspect-ratio-preserving resize strategies (padding, crop, reflect) and the

24 May 2020 • 3 min read

Thermal Infrared Dataset for Object Detection

SUMMARY Thermal infrared cameras detect surface temperature rather than reflected light, which makes them useful for detecting people and objects in darkness, smoke, or other low-visibility conditions where RGB cameras fail. Roboflow released a free 203-image thermal dataset of people and dogs, captured with a Seek Compact XR

13 May 2020 • 7 min read

Data Augmentation in YOLOv4

Learn how data augmentation is used in training YOLOv4 computer vision models.

5 May 2020 • 7 min read

Using Computer Vision to Count Fish Populations (and Monitor Environmental Health)

Counting salmon passing through a fish ladder viewing window is currently done by trained human observers, but object detection models can automate and scale that work. See how data scientist Jamie Shaffer trained a computer vision model.

4 May 2020 • 6 min read

Breaking Down the Technology Behind Self-Driving Cars

SUMMARY Self-driving systems fail at edge cases like a white truck against a bright sky because object detection models learn only from the labeled examples they are trained on, making data collection and annotation the most critical and expensive part of building autonomous perception. This post walks through how

29 Apr 2020 • 4 min read

Breaking Down Roboflow's Health Check Dimension Insights

SUMMARY Resizing images for computer vision is not a simple stretch-to-fill operation: aspect ratio distortion, small image outliers, and extreme width-to-height ratios all affect what a model can learn. Roboflow's Dimension Insights tool, part of the Dataset Health Check, visualizes every image in a

28 Apr 2020 • 1 min read

Introducing the Roboflow Model Library

SUMMARY The Roboflow Model Library brings together 7 open source object detection models and 3 classification models in one place, each accessible via GitHub, a downloadable Jupyter notebook, or Google Colab with free GPU access. All notebooks integrate directly with Roboflow, so datasets uploaded and annotated on the platform can

16 Apr 2020 • 2 min read

Roboflow Presents at Open Data Science Conference (ODSC) East 2020

SUMMARY At ODSC East 2020, Roboflow delivered a three-hour online tutorial walking through the full object detection workflow: framing a good problem, collecting and labeling images, preprocessing, selecting augmentations, training a model, and running inference. Slides, a Colab notebook, and additional resources from the session are publicly available for

10 Apr 2020 • 3 min read

Introducing an Improved Hard Hat Dataset for Computer Vision in Workplace Safety

SUMMARY Detecting hard hat compliance with computer vision requires a well-annotated dataset that reflects the variety of real construction environments. This post introduces a cleaned version of the Northeastern University hard hat dataset, now available on Roboflow Universe with 7,035 images and 27,039 bounding box annotations across

8 Apr 2020 • 1 min read

Our First Video Tutorial: YOLOv3 in PyTorch on a Custom Dataset

SUMMARY This post announces the launch of Roboflow's YouTube channel with a debut video tutorial on training YOLOv3 in PyTorch on a custom dataset, using an Ultralytics implementation and Roboflow for data preparation. The video format was introduced to provide more guided support as the volume and variety

3 Apr 2020 • 2 min read

Using Computer Vision to Fight Coronavirus (COVID-19)

SUMMARY Chest X-rays and CT scans show patterns associated with COVID-19 that computer vision models can learn to flag, offering a supplementary diagnostic path in settings where PCR testing infrastructure is limited. Early open datasets combined COVID-19 scans with existing pneumonia radiology data to give models enough

1 Apr 2020 • 2 min read

Releasing a New YOLOv3 Implementation in PyTorch

SUMMARY Roboflow released a PyTorch implementation of YOLOv3 built on the Ultralytics codebase, complementing the existing Keras version that users had requested an alternative to. In testing on a 12-class chess piece detection task, 300 epochs on Google Colab produced 0.93 recall and 0.978 mAP@50 in

30 Mar 2020 • 1 min read

Introducing Image Preprocessing and Augmentation Previews

SUMMARY Choosing augmentation settings without previewing the result means iterating blind, generating dataset versions only to discover the brightness shift was too aggressive or the rotation too extreme. Roboflow added preprocessing and augmentation previews so you can hover over an image in the dataset builder and see exactly how each

18 Mar 2020 • 3 min read

Introducing Bounding Box Level Augmentations

SUMMARY Standard image augmentation applies transforms to the entire frame, but bounding box level augmentation applies transforms only to the pixels inside each labeled bounding box, letting you vary brightness, blur, or other properties of a target object independently from its background. This technique, drawn from a 2019 Google research

27 Feb 2020 • 4 min read

How to Select the Right Computer Vision Model Architecture

SUMMARY Choosing a computer vision model architecture for object detection involves trading off speed, accuracy, and deployment constraints. This post compares three architectures: YOLOv3, a single-shot detector optimized for speed at 60+ FPS with lower accuracy; MobileNet SSD, a single-shot detector slightly slower than YOLOv3 but stronger on

20 Feb 2020 • 3 min read

How to Win Pioneer

SUMMARY Roboflow held the number one spot on the Pioneer.app global leaderboard for 18 consecutive weeks by following a consistent discipline: set concrete, completable goals each Sunday (not open-ended ones like "continue X"), then exceed them by focusing on controllable inputs rather than unpredictable outputs like

14 Feb 2020 • 2 min read

Releasing an Improved Blood Count and Cell Detection (BCCD) Dataset

SUMMARY The Blood Count and Cell Detection (BCCD) dataset, a widely used object detection benchmark with 364 microscope images across three classes (WBC, RBC, and Platelets), had significant annotation gaps in its original release. The Roboflow team reviewed each image and added 187 missing labels, 183 of them red blood