Model Training

19 Jul 2023 • 5 min read

Improve Accuracy: Polygon Annotations for Object Detection

SUMMARY Polygon annotations consistently outperform bounding boxes for object detection when objects have irregular shapes, because polygons eliminate background noise that bounding boxes include. An experiment comparing both annotation types, combined with augmentations (rotation, saturation, cutout, bounding box shear) and pretrained weight initialization, showed meaningful mAP gains at each step.

19 Jul 2023 • 9 min read

Pose Estimation Algorithms: History and Evolution

SUMMARY Pose estimation, the computer vision task of detecting key body joints in images and video, has progressed from hand-crafted geometric methods in the 1990s through graphical models in the 2000s to convolutional neural network approaches that now dominate the field. This overview traces that arc, covering traditional methods,

17 Jul 2023 • 5 min read

How to Build a Semantic Image Search Engine with Supabase and OpenAI CLIP

SUMMARY OpenAI's CLIP model encodes both images and text into a shared embedding space, making it possible to search an image library with a plain text query and get back visually and semantically relevant results. This tutorial builds that system by calling the CLIP model through Roboflow'

5 Jul 2023 • 4 min read

Detecting Objects with DETIC vs Custom Training

Learn how to evaluate large foundation models and how custom model training can improve performance.

23 Jun 2023 • 3 min read

CVPR 2023 Highlights

Three members of the Roboflow team attended CVPR this year. Read our highlights from the conference and what trends we noticed.

19 Jun 2023 • 4 min read

How to Train DETR on a Custom Dataset

Learn how to train a DETR model on a custom dataset.

16 Jun 2023 • 9 min read

What is Hyperparameter Tuning? A Deep Dive

This guide explores what hyperparameter tuning is, common hyperparameters in computer vision, methods of tuning hyperparameters, and more.

13 Jun 2023 • 5 min read

Launch: Evaluate Computer Vision Models on Roboflow

In this guide, we walk through how to evaluate computer vision models hosted on Roboflow using our in-app model evaluation tool.

30 May 2023 • 6 min read

How to Train a YOLOv8 Classification Model

In this guide, we walk through how to train a classification model using YOLOv8 and a dataset hosted on Roboflow.

16 May 2023 • 7 min read

How to Train YOLO-NAS on a Custom Dataset

YOLO-NAS is the latest state-of-the-art real-time object detection model. Learn how to train YOLO-NAS on your custom data.

10 May 2023 • 12 min read

Multimodal Models and Computer Vision: A Deep Dive

In this post, we discuss what multimodals are, how they work, and their impact on solving computer vision problems.

10 May 2023 • 5 min read

How to Evaluate Computer Vision Models with CVevals

Learn how to calculate precision, recall, f1, and confusion matrices for your computer vision models with CVevals.

1 May 2023 • 6 min read

Leveraging Embeddings and Clustering Techniques in Computer Vision

Explore the world of image embeddings in computer vision, as we dive into clustering, dataset assessment, and detecting image duplication. Discover dimensionality reduction techniques like t-SNE and UMAP. Use CLIP embeddings for analyzing image class distribution and identifying similar images.

21 Apr 2023 • 5 min read

Zero-Shot Image Annotation with Grounding DINO and SAM - A Notebook Tutorial

In this comprehensive tutorial, discover how to speed up your image annotation process using Grounding DINO and Segment Anything Model. Learn how to convert object detection datasets into instance segmentation datasets, and use these models to automatically annotate your images.

7 Apr 2023 • 6 min read

What is Segment Anything Model (SAM)? A Breakdown.

Learn how Meta Research's new Segment Anything Model works to achieve high performance on image segmentation tasks.

30 Mar 2023 • 6 min read

Grounding DINO : SOTA Zero-Shot Object Detection

Most object detection models are trained to identify a narrow predetermined collection of classes. Zero-shot detectors like Grounding DINO want to break this status quo by making it possible to detect new objects without re-training a model.

16 Mar 2023 • 10 min read

Speculating on How GPT-4 Changes Computer Vision

OpenAI released GPT-4 showcasing strong multi-modal general AI capabilities in addition to impressive logical reasoning capability. Are general models going to obviate the need to label images and train models?

22 Jul 2022 • 9 min read

How To Train SegFormer on a Custom Dataset

In this post, we will walk through how to train SegFormer on a custom dataset using Pytorch Lightning to classify every pixel in an image.

13 Jul 2022 • 6 min read

How to Train a YOLOv7 Model on a Custom Dataset

SUMMARY This tutorial walks through training a YOLOv7 object detection model on a custom dataset using a Google Colab notebook, covering dependency installation, loading a dataset from Roboflow Universe in YOLOv7 format, running training, evaluating results with detect.py, and deploying the model via API or to edge devices like

24 Jun 2022 • 4 min read

CVPR 2022 - Best Papers and Highlights

In this post, we take the opportunity to reflect on the computer vision research landscape at CVPR 2022 and highlight our favorite research papers and themes.

15 Jun 2022 • 9 min read

Train and Deploy YOLOS Transformer On a Custom Dataset

In this post, we showcase training and deploying YOLOS end to end, from labeling your data, to training your model, to deploying your model on AWS for inference.

19 May 2022 • 2 min read

Launch: Instance Segmentation Project Training and Inference

SUMMARY Roboflow now supports instance segmentation project training and inference natively on the platform, letting teams train directly from a pre-trained COCO checkpoint without writing training code. Instance segmentation produces per-object outlines rather than just bounding boxes, making it the right choice when downstream tasks require measuring object

13 Apr 2022 • 6 min read

How to Train Detectron2 for Custom Instance Segmentation

A walk through on how to train Detectron2 to segment your custom objects from any image by providing our model with example training data.

9 Mar 2022 • 3 min read

Introducing New and Improved Roboflow Train

SUMMARY Roboflow Train now lets users choose between Fast and Accurate model types, giving engineers direct control over the speed-versus-accuracy tradeoff at training time. New users receive three Fast training credits to test the managed training and deployment pipeline at no cost. Most custom tasks see improved accuracy