TelecomNet: The ImageNet of the telecom world

Inspired by the ImageNet, TelecomNet – initiated this year by Ericsson – will help to significantly accelerate the development of computer vision applications within the telecom industry. Read below to find out how.

Engineer working in a server room
Sep 04, 2019
Vik Li 

Investment Director

Volodya Grancharov 

Principal Researcher Digital Representation and Interaction

Manish Sonal  

Experienced Researcher Digital Representation and Interaction

David Hunter  

Strategic Product Manager Automation

Category, topic & hashtags

The idea for TelecomNet first came about during the development of a use case which automated tower inventories from drone images. Our first instinct, naturally, was to look to the Internet. There are several open datasets of labelled images, such as ImageNet, but none of these datasets has enough telecom-related images to meet our needs.

Our next instinct was to look closer to home. At Ericsson, there are multiple teams who have painstakingly labelled tens of thousands of telecom-related images for their own applications. However, these labelled images were not easy to share because they used different naming conventions, different label formats, and were stored in different places.

What is ImageNet?

ImageNet project, started by Fei-Fei Li from Stanford in 2006, has over 14 million images with annotation of everyday life objects.  The famous ImageNet Large Scale Visual Recognition Challenge is a catalyst for great progress in computer vision with algorithms that surpass the human-level performance.

What is TelecomNet?

Initiated this year by Ericsson, TelecomNet is a crowd-sourced computer vision database, similar to ImageNet, developed specifically for the telecom world.  Based on Ericsson’s extensive knowledge and data about all the telecom equipment and telecom sites, TelecomNet provides functions such as crowd sourcing of images, video/image annotations with the EVA video annotator, and searching, exploring & downloading of images and labels. 

Automation supported by computer vision

The advance of computer vision technology continues to be of great help in automating many business processes across telecom. A couple of examples include the automatic generation of inventory for telecom sites’ equipment, and accelerated customer acceptance through a 3D model of the cell site with accurate dimensions and recognized equipment. In addition, computer vision technology provides automation and support for the field technician, automatic detection of abnormalities such as material corrosion, incorrectly connected cables or tilted antennas.

This automation is increasingly important with the rollout of 5G networks. Complexity of the site rollout task and the number of sites require digitalization and automated visual inspection of the installed components.

All such computer vision applications require annotated image data – and a lot of it. For computer vision researchers and developers, TelecomNet removes the operational burden to annotate and maintain visual data, allowing them to focus on the productive computer vision task. With TelecomNet, we aim to dramatically accelerate the development of different computer vision applications for the 5G network rollout.

TelecomNet architecture

TelecomNet receives heterogeneous visual data of different telecom objects or scenarios. These data are from multiple sources such as site inspection with drones, field technicians capturing site conditions with a handheld camera, or other visual data from special equipment such as lidar, thermal cameras etc.

The core of TelecomNet is the structured visual data and the corresponding annotations, made available to any team in Ericsson that are developing computer vision applications. The high-level diagram of TelecomNet is depicted in Figure 1.

Figure 1. TelecomNet consists of four main modules: Annotated Image Repository, GDPR Pre-Processor, Distributed storage, and Video Annotator (EVA).

Figure 1. TelecomNet consists of four main modules: Annotated Image Repository, GDPR Pre-Processor, Distributed storage, and Video Annotator (EVA).

Annotated image repository

All images are stored in file systems, with image information and annotation data stored in the database. The raw visual data is stored and made searchable by related metadata on TelecomNet.  Like Wordnet being used by ImageNet, we have constructed an Ericsson Product Family Tree label set largely following our product portfolio structure, so the users can explore the different categories of images. There are also label sets that describe different scenarios of telecom use cases, such as correct or incorrect installation of certain components, conditions of the telecom sites etc. The label files are generated dynamically and can be downloaded by the users together with the original image files. Since the annotation data are stored in the database, we can generate different label formats according to users’ needs. 

The GDPR pre-processor

The platform crowd-sourced different visual data that do not contain any sensitive customer or personal information. Although we take every means possible to avoid uploading any images that contain personal information, we have also implemented a simple solution that

can identify if there are any human faces, car registrations or other information that can identify a person contained within the uploaded images. Those images will then be deleted from the platform and the users who uploaded those images will be notified.

Distributed storage

Where local jurisdiction dictates that image files need to be kept physically within a country’s border, we implement a distributed storage solution that spans different geographic locations with restricted access to those images.

EVA video annotator

The EVA video annotator allows the TelecomNet platform to efficiently label large amounts of visual data. It is designed on a client server architecture and implemented in Django framework, see Figure 2. Its integrated tracker minimizes manual intervention and accelerates the annotation process. It outputs annotations in YOLO or Pascal VOC format, which makes them suitable for direct use in a Machine Learning algorithm optimizing an object detector.

Figure 2. EVA* architecture, backend system is based on Django Framework with Celery and Redis handling asynchronous tracking related tasks.

Figure 2. EVA* architecture, backend system is based on Django Framework with Celery and Redis handling asynchronous tracking related tasks.

The vision

The immediate next step is to implement APIs for the platform to either receive raw visual data and other annotated images, or to send annotated images to other Machine Learning design or execution platforms. Further, the current metadata or label-based search capabilities will be enhanced with search in video or large image sets by a visual sample.

Like ImageNet which has stimulated many great computer vision algorithms, the biggest ambition for TelecomNet is to rapidly accelerate the progress of computer vision applications within the telecom industry, and all subsequent automation that comes with it.


The Ericsson Blog

Like what you’re reading? Please sign up for email updates on your favorite topics.

Subscribe now

At the Ericsson Blog, we provide insight to make complex ideas on technology, innovation and business simple.