About Portal

The FathomNet Portal helps users train their own machine learning models by providing intuitive interfaces for annotation, streamlined access to the necessary computational power, and accessible tools to analyze the output.

Overview

Within the FathomNet Portal, users can upload and view video, run machine learning algorithms to automatically identify and classify organisms (or do it manually themselves), and search for observations recorded by others. The customized privacy, and user roles allowed for a tailor experience for each Portal project.

Glossary

Localization/Annotation  - Describes a region of interest in a frame of video, or a still image. Localizations in portal are box-shaped described by an x, and y position, and height and width.

Track/Observation - When a localization of an animal for example is shown moving across a region in a video, the same animal can be boxed in a series of frames. When localizations are made and associated together, they can individually be referred to as a track. 

Related Data - Usually sensor readings like Oxygen, Water Temperature, etc. This data can be submitted after media upload. Submitted sensor data is shown in the player view, and included in export.

QA/QC - Shorthand for Quality Assurance or Quality Control which is a common process in data annotation, and is an important part of training algorithms as well. In portal, QA/QC flags can be set at the media level by Editors such as “Ready for QA” and at the Observation level by Verifiers, and up.

Permissions - Define what actions can be taken within a project. Groups of permissions are set based on roles. Default permissions in a project are first set by choosing the privacy level of the project as a whole.

Projects - User-created groups that can be managed granularly by permission on the project, and by assigning roles to users on the team.

Roles - In portal there are a set user roles and their level of access to a projects media and observations.

Folders - Can be created within a project, or another folder to structure content or partition media for a QA/QC workflow. Folders inherit the permissions of the parent project.

Algorithm - Refers to the combined toolage we use to generate information with less human input. Portal algorithms use fine-tuned metrics from specific Models to determine regions of interests, and add a Label if it is suggested.

Weights - Training data creates fine tuning for specific use cases for each named algorithm in our system. We currently offer two MBARI-trained weights added via a model submission form on the FathomNet Portal website by ML Ops team.

Model - Can be used interchangeably with algorithm. Multiple models which have specific use cases and strengths.

Detection - A region of interest in a video frame or image. Algorithms may be “detection only” meaning all results will have “No Label” or be generically labelled as “object.”

Classification - Assignment of a label to an Observation by an algorithm. Algorithms may be “classification only” meaning they suggest the content, but won’t box it. A trained classification and detection algorithm is ideal in that it does both.