February 2023 Summaries

The UCF101 Action Recognition dataset is a collection of realistic action videos with 101 action categories, curated from YouTube. It was the largest video action recognition dataset available at its release in 2012 and covers various computer vision use cases such as video, images, location, point-cloud, action-recognition, classification, detection, segmentation, and relationships. The dataset is characterized by a large diversity of actions and variations in camera motion, object appearance and pose, scale, viewpoint, background, and illumination conditions. It has been used for human action recognition tasks with real-world applications like surveillance, healthcare, human-computer interaction, content-based video retrieval, and video summarization. The dataset is available as a part of the FiftyOne Dataset Zoo, which can be accessed using simple commands in the FiftyOne toolset.

Voxel51, a company focused on bringing transparency and clarity to the world's data through their open source FiftyOne computer vision toolset, is currently hiring for remote positions. The company aims to create a diverse team of exceptional individuals who are passionate about contributing to the field of data-centric machine learning. In this series of posts, readers can get an inside look at the work and lives of Voxel51's team members. This week, Software Engineer Lanny Wang shares her experience working on improving features for FiftyOne and highlights her favorite project so far: "Search by text similarity." She enjoys the independence and collaboration within her role and finds fulfillment in contributing to open source projects. When not working, Lanny loves caring for her tropical garden and is planning to grow mushrooms in the next year. Voxel51 currently has openings for a VP Engineering and DevOps Engineer.

This week's FiftyOne tips and tricks blog covers various topics such as omitting classes with few detection instances, saving changes to sample fields, predicting class labels in homogeneous images, matching classification results, shutting down a session, and more. The blog provides detailed explanations and code examples for each topic, helping users understand how to effectively use FiftyOne's open-source machine learning toolset for computer vision models.

This article is the third and final part of a series that demonstrates how to work with YOLOv8, from downloading off-the-shelf models to fine-tuning these models for specific use cases. The author uses two libraries, FiftyOne and Ultralytics, to fine-tune a YOLOv8 model for custom computer vision applications. The article covers the process of defining a use case, choosing training data, fine-tuning the YOLOv8 model, and assessing the improvement in performance. It also discusses potential approaches to further improve the model's performance.

In this article, the author continues a three-part series on YOLOv8, focusing on evaluating model predictions. The evaluation process includes printing performance metrics, viewing concerning classes, and finding poorly performing samples. The author uses FiftyOne, an open-source computer vision toolkit, to analyze the quality of YOLOv8n detection model's predictions. They demonstrate how to evaluate a YOLOv8 model's performance on specific datasets and provide insights into improving the model's effectiveness for custom applications.

This article is the first in a three-part series that explores YOLOv8, a state-of-the-art object detection model. In this part, you will learn how to generate, load, and visualize YOLOv8 predictions using FiftyOne, an open-source computer vision toolkit, and Ultralytics, the library that provides access to YOLOv8. The series is organized into three parts: generating and loading YOLOv8 predictions, evaluating YOLOv8 model predictions, and fine-tuning YOLOv8 models for custom computer vision applications. The article begins with a brief history of YOLO models since their initial release in 2015 and how they have evolved over time to become one of the most popular object detection models in the field. It then introduces YOLOv8, which was released by Ultralytics in late 2022 and comes with a new backbone that includes general-purpose models for object detection and instance segmentation tasks. To get started, you need to install the Ultralytics and FiftyOne Python packages and import relevant modules from both libraries. The article demonstrates how to apply YOLOv8 models to individual images or a set of images in a directory using the command line interface provided by Ultralytics. It also explains how to load your images into a FiftyOne Dataset and export them in YOLOv5Dataset format, which is compatible with YOLOv8. The article then focuses on visualizing YOLOv8 predictions using the FiftyOne App. It demonstrates how to generate detection predictions for a subset of the MS COCO dataset from the command line and add these predictions to the dataset in FiftyOne's Detections format. The article also provides functions to read YOLOv8 detection prediction files, convert them into FiftyOne Detections objects, and efficiently add these detections to all samples in a dataset by batching the read and write operations to the underlying MongoDB database. Finally, the article concludes with a brief overview of how to visualize YOLOv8 predictions on images using the FiftyOne App and how to convert YOLOv8 instance segmentations to FiftyOne instance segmentations. In Part 2 of this series, you will learn how to evaluate the quality of YOLOv8 model predictions, identify edge cases, and assess potential modes of failure.

Voxel51, an open source computer vision toolset developer, has released the latest version of its product, FiftyOne 0.19. The company is committed to transparency and giving back to the computer vision community through sponsoring Meetups, contributing to charitable causes on behalf of the community, and developing open source tools like FiftyOne. Community members are using FiftyOne in various sectors such as research and development, agriculture, waste management, and materials recovery. The new release includes features like a customizable Spaces framework, in-App embeddings visualization, saved views, on-disk segmentations, and more.

This week's FiftyOne tips and tricks blog covers adding and merging data. The text introduces FiftyOne, an open source machine learning toolset that helps improve the performance of computer vision models by enabling high-quality dataset curation, model evaluation, error detection, embedding visualization, and faster deployment to production. It then delves into various methods for adding and merging data in FiftyOne, including handling samples appearing multiple times, using add_dir() and merge_dir() methods for datasets stored in directories or archives, and adding model predictions with the merge_samples() method. The text also discusses exporting multiple labels with merge_labels() and joining the FiftyOne community on Slack, GitHub, Meetup, and other platforms.

Voxel51 and the FiftyOne community have released version 0.19 of their open source machine learning toolset, which includes new features to improve computer vision datasets visualization and boost model performance. The release introduces Spaces, an all-new customizable framework for organizing interactive information panels within the FiftyOne App; in-app embeddings visualization; saved views; on-disk segmentations; and enhanced UI filtering options. Additionally, documentation for FiftyOne Teams is now publicly available.

Voxel51 hosted the February 2023 Computer Vision Meetup featuring Paula Ramos and Vishal Rajput as speakers. The event covered topics such as Anomalib, OpenVINO, Intel Geti, and OpenAI's Whisper Model for speech recognition. Anomalib is a library for unsupervised anomaly detection from data collection to deployment, while OpenVINO is an ecosystem that includes tools like Anomalib and the Intel Geti platform. The Whisper Model by OpenAI uses weak supervision to scale its data and improve accuracy in speech recognition tasks. The Computer Vision Meetup network has grown to nearly 3,000 members across 13 locations worldwide, with upcoming events scheduled for March, April, and May featuring various speakers discussing computer vision topics.

This week's FiftyOne tips and tricks blog covers various topics such as isolating spurious or missing objects, filtering by ID, merging datasets with shared media files, exporting GeoJSON annotations, and picking random frames from videos. The blog also encourages users to join the FiftyOne community for further assistance and collaboration.

The article discusses a template created by Tarmily Wen, Principal Computer Vision Research Engineer at ADT, using cookiecutter to automate the setup process for each new machine learning (ML) project. This approach presents several benefits including consistency, reproducibility, data standardization, and speed. The python ML project cookiecutter template adds essential features such as environment reproducibility, code consistency, and data standardization, along with some "nice to haves" like documentation generation and test automation to every ML project. It uses docker for a consistent environment inside of the container, poetry for consistent Python package dependencies, and FiftyOne for data visualization and formatting. The template also provides continuous integration (CI) using GitHub Actions to maintain code quality.

This week's FiftyOne tips and tricks blog delves into model evaluation. The open source machine learning toolset, FiftyOne, enables data science teams to improve the performance of their computer vision models by helping them curate high quality datasets, evaluate models, find mistakes, visualize embeddings, and get to production faster. Model evaluation methods in FiftyOne include regressions, classifications, detections, polygons, instance and semantic segmentations, on both image and video datasets. The Evaluation API supports common computer vision tasks like object detection and classification with default evaluation methods that implement some of the standard routines in the field.