Custom object detection with Customvision.ai and using trained model on Android device

Computer Vision from the Cognitive service is the way to find meaning in the image, but what to do when we have to find our own custom object?
How to create our own model to find our custom object?

If you are looking for a simplest way, then Microsoft customvision.ai is the one to look for.
Just head over to below link & register yourself.

https://customvision.ai/




The Custom Vision Service is an Azure Cognitive Service that lets you build custom image classifiers. It makes it easy and fast to build, deploy, and improve an image classifier.

Once you are registered with custom vision, you can create 2 projects for free.
Create a new project. Use "Classification" as Project Type, "MultiClass (Single tag per image)" as Classification type (I am using images with only single object in it) and "General (Compact)" as Domain for leaner model for Mobile device use.
Refer below image for reference.


Now upload the images & provide a tag like "londonBusCoffeeMug".


Next step is to train our model for the images.
Just hit the Train button on the top.
Here is the training performance. (impressive to see 100% Precision & 100% Recall 😏)
We can do a quick prediction using the Predictions tab, or export the model to see in action in a mobile app. I will do the later here.
Customvision give option to export the trained model in CoreML, TensorFlow or ONNX format.



I will export to TensorFlow to use in the Android App. The exported tensorflow model contains model.pb & labels.txt, the Model & Labels respectively.

Now we have our model ready, so lets test it on Android device. (I expect you have android studio configured on your machine)
Here is a sample app from Microsoft where we can put our model & test.

https://github.com/Azure-Samples/cognitive-services-android-customvision-sample

Or do a git clone from https://github.com/Azure-Samples/cognitive-services-android-customvision-sample.git

Once you have got the sample application cloned, just replace the model.pb & labels.txt from the apps assets folder.
Now just build & run the application on a device.

If you give some image which resembles to the object we trained our model on, then it should give proper object name in the pink box as below.


Hope this help in creating your custom object detection & testing on device.


Happy Coding,
Parshuram

Comments

Unknown said…
Great walkthrough!!!
2:56 PM
Parshuram said…
Thanks Pankaj.
7:49 PM
Post a Comment

Popular posts from this blog

6000 KM Road Trip

Image
Embarking on a road trip had always been our goal since we acquired our first car back in 2009. However, between demanding work schedules and challenging road conditions, the idea remained a distant dream. It wasn't until 2023 that circumstances shifted, granting us the freedom to prioritize exploration over work commitments. In January 2023, we finally indulged in a short but exhilarating road trip across Karnataka, spanning over 2000 kilometers. Despite the distance, the experiences we encountered along the way were nothing short of extraordinary. Inspired by our memorable journey, we resolved to delve deeper into the rich tapestry of India's cultural and historical treasures. Our itinerary took shape, beginning with the awe-inspiring Ajantha and Ellora caves in Aurangabad, then venturing to the sacred sites of Puri and Konark along the eastern coast. From there, our route traced a path southward through Andhra Pradesh and Tamil Nadu, culminating at the southernmost tip
Read more

Export MS Access report to Excel..Font Problem

I was working on a small assignment on access and trying to export an Access report to Excel. Everything worked when until i opened the exported Excel, to my surprise font colour of all the lines in the excel was very light. I tried to change the field color, font weight and all possible properties, nothing seems to be working. But finally i got the solution. The trick was 1. Make the back style of the the fields to "Transparent" which was default "Normal" 2. Make the Fore Color as Automatic black from the eclipse next to the property. Spent few hours trying to make things work so sharing it for the people. Happy Coding, Parshuram
Read more

Testing AWS Rekognition from POSTMAN

Image
I was trying AWS Rekognition & before using it in the app, i wanted to test with POSTMAN. But unfortunately there are not many resources available to be useful. Thus showing the simple way to setup POSTMAN. Obviously before jumping to POSTMAN, we need to have an AWS account setup and a user been created. (The user should have Rekognition related permissions) Also make sure to note the Access Key & Secret Key.  Lets start the POSTMAN & create a new POST request. The url to hit is  https://rekognition.ap-south-1.amazonaws.com  . It could change depending on your selected AWS region, mine is ap-south-1. First thing to set is the Authorization & select "AWS Signature" from the Type. Enter AccessKey, SecretKey which we noted while creating the new user. AWS Region should be from url & Service Name is "rekognition". We need to add 2 extra headers, "Content-Type" as "application/x-amz-jon-1.1" & "X-Amz-
Read more