Azure Custom Vision - Object Detection

Azure Cognitive Services Custom Vision using the JavaScript SDK for Object Detection.

Infrastructure

Create the Cognitive Services resource:

az cognitiveservices account create -n '<name>' -g '<group>' --kind CognitiveServices --sku S0 -l 'eastus2' --yes

For Cognitive Services --sku F0 tier is no longer available.

Project Setup

Start by copying the sample .env:

cp example.env .env

Go to customvision.ai resources and copy the folling parameters:

💡 I’ve opened #14595 to get the API Key using the CLI.

Now add the values to the .env file:

customVisionTrainingKey='<training_key>'
customVisionTrainingEndPoint='<endpoint>'
predictionResourceId='<prediction_resource_id>'

ℹ️ Custom Vision recommends at least 50 images per set to ensure model performance.
Following the rule of thumb 70/30 you should have at least 15 additional images for the prediction tests.

Set the remaining configuration for your project:

# Project 
projectName='<your_project_name>'
publishName='<publish_name>'
# Your sample data
trainingSampleDataRoot='<path>'
predictionSampleDataRoot='<path>'
# Separeted by commas (`,`):
tags='tag1,tag2,tag3'

Training and Prediction

If you haven’t already, download the dependencies:

yarn install

First you need to create your project and tags:

# This will create and output the project id
ts-node src/createProject.ts

Add the project ID output to the .env file:

projectId='xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx'

Upload your images (my script is limited to 64 images - help me with a pull request 😁):

ts-node src/upload.ts

My algorithm is limited to send a single batch of 64 images - help me with a pull request 😁

Now, go to the Custom Vision Portal and tag the images. You may add multiple tags for the same image.

You need to have at least 15 items for each tag for object detection.

After you tagged all the images, train your model. This will take a while:

ts-node src/training.ts

Run the prediction tests:

ts-node src/prediction.ts

This will also publish the iteration. You must unpublish previous iterations on the Performance tab at customvision.ai if you wish to publish new ones.

You’ll get an output for each image:

Results:
         budweiser: 1.23% 0.00067794323,0.0699334145,0.038740702,0.9300566
         budweiser: 96.32% 0.182365745,0.24337104,0.104204118,0.756619
         budweiser: 7.03% 0.2837638,0.269703448,0.09547615,0.718327045
         budweiser: 89.82% 0.358226,0.267752945,0.0972617,0.724578857
         budweiser: 96.16% 0.45870012,0.224418223,0.111133575,0.7673404
         budweiser: 94.66% 0.572847,0.299463362,0.08857638,0.6836401
         budweiser: 2.72% 0.637388051,0.405668736,0.130907238,0.464525819
         budweiser: 47.84% 0.728372,0.4846303,0.09750664,0.4907381
         budweiser: 2.77% 0.27285856,0.62536025,0.108680427,0.330352247
         budweiser: 2.41% 0.6389212,0.652789831,0.104773819,0.2945472
         budweiser: 6.49% 0.153952986,0,0.193375885,0.9743123
         budweiser: 2.16% 0.419736058,0,0.207257,0.963360548
         budweiser: 1.18% 0.643069267,0.09114501,0.2224654,0.905299067
         budweiser: 4.31% 0.226873457,0.275986224,0.226002991,0.7033454
         corona: 2.04% 0.5713419,0.3917932,0.0912797451,0.5812621
         corona: 10.40% 0.728372,0.4846303,0.09750664,0.4907381
         corona: 1.93% 0.6389212,0.652789831,0.104773819,0.2945472
         heineken: 1.53% 0.45870012,0.224418223,0.111133575,0.7673404
         heineken: 3.57% 0.5713419,0.3917932,0.0912797451,0.5812621
         heineken: 3.46% 0.728372,0.4846303,0.09750664,0.4907381

You may go to the portal now and see the results:

References

Custom Vision with SDK

Custom Vision Service Limits and Quotas